knowledge-review

name	knowledge-review
version	1.1.0
description	[Research] Use when you need to review knowledge artifacts for completeness, citation quality, confidence accuracy, and template compliance.

Quick Summary

Goal: Ensure knowledge artifacts are evidence-backed, complete, protocol-compliant, and safe to use for decisions — reviewing for quality, completeness, citation accuracy, and template compliance.

Summary:

• PURPOSE: READ-ONLY audit of a knowledge artifact (research report / course / strategy) for completeness, citation accuracy, confidence calibration, source quality, and template compliance — output is a verdict, NEVER an edit.
• MAIN STEPS in order: (1) read the artifact → (2) run the 7-checklist audit — template compliance · citation audit · confidence accuracy · source quality · knowledge gaps · cross-validation · actionability, verifying presence AND quality depth (never just that a section exists) → (3) run the adversarial Anti-Bias Gate → (4) emit PASS/WARN/FAIL per-check + verdict (APPROVED/REVISE/BLOCKED) → (5) conditional Round 2 focused re-review.
• Default to SKEPTIC: run the Anti-Bias Gate before any verdict — find a contradicting source per major claim, stress-test every score ≥80%, state the strongest alternative conclusion, check supporting-vs-contradicting source ratio, run a pre-mortem, and argue the opposite verdict in 2+ sentences.
• Calibrate confidence to evidence: a single source ≠ 80%, scores >80% need 2+ independent sources with contradicting evidence addressed, single-source claims marked unverified must be <60%, and findings <60% must be flagged prominently.
• Convergence: a clean Round 1 ENDS the review; any finding triggers validate → fix → full re-review until zero issues.

Workflow:

1. Read artifact — Load the knowledge report/course/strategy
2. Template compliance — Verify all enforced sections present
3. Citation audit — Check inline citations and source table
4. Confidence check — Verify scores match evidence
5. Output review — Summary with pass/warn/fail per check

Key Rules:

• Every section from template must be present and non-empty
• Every factual claim must have inline citation
• Confidence scores must match evidence basis
• READ-ONLY — do not modify the artifact

Be skeptical. Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence percentages (Idea should be more than 80%).

First Principle — Easy to Change

Success metric of every coding decision: future change cost. DRY, SRP, abstraction, design patterns, naming, layering, tests — every technique serves one goal: making next change cheaper.

When evaluating code, refactor, test, or abstraction, ask: does this make next change cheaper or more expensive?

• Reject "best practices" raising change cost (premature abstraction, speculative generality, leaky indirection, ceremony without payoff).
• Name real enemies in findings: coupling, hidden state, duplicated knowledge, unclear intent, irreversible decisions exposed too early.
• Simpler design that is easy to change beats sophisticated design that isn't.

Apply this lens before invoking any specific rule, pattern, or checklist below — if downstream rule would raise change cost, this principle wins.

---

Adversarial Review Mindset (NON-NEGOTIABLE)

Default stance: SKEPTIC challenging research quality, NOT confirming research completeness.

Source confirmation bias trap: AI gravitates toward sources confirming its working hypothesis. Knowledge artifact built iteratively — by completion, framing is locked in. This section forces challenge of both sources AND framing.

Adversarial Techniques (apply ALL before concluding)

1. Source Bias Detection Top 3 claims: "What sources CONTRADICT this claim?" No contradicting source cited → either reviewer didn't look, or evidence truly one-sided. Ask: "What would skeptic of this conclusion cite?" No counterevidence addressed → confidence score inflated.

2. Confidence Calibration Challenge Each confidence score ≥ 80%: "What would need to be true for this confidence to be wrong?" High confidence warranted ONLY when: (a) multiple independent sources agree, (b) contradicting evidence addressed, (c) methodology sound. Challenge any score resting on single source or undisclosed assumptions.

3. Alternative Conclusion Check Given same evidence, what DIFFERENT conclusion could reasonable expert reach? Artifact not addressing 1+ credible alternative interpretation → analysis incomplete. State strongest alternative conclusion.

4. Cherry-Picking Detection Count sources supporting main conclusion vs. sources challenging it. Ratio > 3:1 favoring supporting sources without explicit explanation of why contradicting sources discounted → flag cherry-picking.

5. Pre-Mortem Assume artifact's recommendation implemented and fails. Write most plausible failure scenario given research limitations. Artifact not acknowledging this failure mode → missing a risk section.

6. Contrarian Pass Before writing any verdict, generate 2+ sentences arguing OPPOSITE conclusion about artifact's quality. Then decide which argument is stronger.

Forbidden Patterns

• "Sources are cited" → Presence of citations ≠ quality. Do they actually support the claim?
• "Confidence scores look reasonable" → What would LOWER the confidence score? Name it.
• "Comprehensive coverage" → What perspective is MISSING from this research?
• "Recommendations are actionable" → On what evidence? What's the confidence of the evidence chain?
• Approving a knowledge artifact without challenging the evidence quality → Forbidden.

Anti-Bias Gate (MANDATORY before finalizing verdict) (MUST ATTENTION)

• found 1+ contradicting source per major claim (or flagged its absence)
• challenged 1+ confidence score ≥ 80% with a stress test
• stated strongest alternative conclusion from same evidence
• checked source balance (supporting vs. contradicting ratio)
• ran pre-mortem on main recommendation
• generated 2+ sentences arguing opposite verdict

Any item unmet → adversarial review incomplete. Go back. — why: skipping the gate ships confirmation-biased verdicts as fact.

Knowledge Review

Review Checklist

1. Template Compliance

#	Check	Presence	Quality Depth
1	All enforced sections present — every required section from the template exists in the artifact	Are ALL required sections present (not just most)? Is any section empty vs. placeholder?	Does each section contain substantive content, or is it a heading with nothing beneath it? A partially-filled section is as dangerous as a missing one.
2	No sections are empty placeholders — section bodies contain real content, not "TBD" or "to be filled"	Are section bodies substantive or just "TBD / to be filled"?	Is the content specific to this artifact, or generic filler? A section that says "risks will be identified later" has negative value — it creates false confidence.
3	Section order matches template — sections appear in prescribed sequence	Do sections appear in the prescribed order?	Does reordering break any cross-references between sections? If section 3 references section 2, out-of-order placement creates reading confusion.

2. Citation Audit

#	Check	Presence	Quality Depth
1	Every factual claim has inline citation [N] — all assertions are backed by a numbered source reference	Are ALL claims cited, or only the obvious ones? Uncited claims are assertions, not findings.	Do the cited sources actually say what the text claims? A source cited for a paraphrase is different from a source cited for a direct claim.
2	Every source in Sources table is referenced in text — no source appears in the table without a corresponding [N] reference	Are all table sources referenced in the text body? Orphan sources = sources added for credibility, not used for evidence.	Are sources cited at the most specific claim they support, or cited vaguely at section level? Section-level citation hides which sub-claims are actually supported.
3	No orphan citations — every [N] reference in the text matches a Sources table row	Are there [N] references that don't match any Sources table row?	Are citation numbers consistent throughout (no gaps, no duplicates)? Broken citation numbering signals the artifact was edited without maintaining integrity.
4	Sources table has: Title, URL, Author, Date, Tier — all five fields present for every source	Are ALL 5 fields filled? Is Tier assigned (not just blank)?	Are the Tier assignments accurate? A blog post assigned Tier 1 inflates perceived source quality. Is the Date current enough for the topic?

3. Confidence Accuracy

#	Check	Presence	Quality Depth
1	Per-finding confidence scores declared — each finding has its own explicit confidence percentage	Does EACH finding have its own score, or is one artifact-level score applied everywhere?	Are scores at the finding level, not the section level? A single score per section hides that some findings within it may be poorly supported.
2	Overall confidence declared — an aggregate confidence score for the artifact is stated	Is the overall score present and explicitly stated?	Is the overall score a reasoned aggregate, or just the highest per-finding score? An average of 85%, 60%, and 40% is NOT 85%.
3	Scores match evidence basis (not inflated) — confidence percentages are calibrated to actual source count and quality	Is each score justified by the number and quality of independent sources? A single source ≠ 80%.	What would LOWER this confidence score? If no answer exists, the score is likely inflated. Are scores above 80% supported by 2+ independent sources with contradicting evidence addressed?
4	Findings <60% flagged prominently — low-confidence findings are visually distinct from high-confidence ones	Are low-confidence findings visually distinct (e.g., ⚠️ prefix), not buried in body text?	Are low-confidence findings positioned to prevent downstream misuse? A low-confidence finding mentioned once in passing will be treated as fact by readers who skim.

4. Source Quality

#	Check	Presence	Quality Depth
1	Tier distribution appropriate (not all Tier 4) — source tiers are spread across quality levels	Is the Tier distribution recorded? Are Tier 4 sources a minority?	Does Tier distribution match the claim importance? A Tier 4 source for a core claim is a risk regardless of how many Tier 1 sources exist elsewhere.
2	At least 50% Tier 1-2 sources for key claims — high-stakes conclusions rest on authoritative sources	Do key claims cite Tier 1-2 sources? Is the 50% threshold met overall?	Are the Tier 1-2 sources actually authoritative for THIS specific claim domain, or just prestigious in a different domain? Domain mismatch inflates perceived authority.
3	Recency appropriate for topic type — sources are current relative to how fast the topic evolves	Are sources dated? Is recency assessed for each source?	Is "recency" calibrated to the topic's rate of change? A 2019 source on cloud pricing is stale; a 2019 source on database theory may be fine. Are any outdated sources flagged?

5. Knowledge Gaps

#	Check	Presence	Quality Depth
1	Gaps section is present and honest — a dedicated section describes what the research did NOT find	Is a Gaps section present? Does it list specific gaps, not just "more research needed"?	Are the gaps specific enough to guide follow-up research? "Unknown pricing" is actionable; "some uncertainty exists" is not. Are gaps ranked by impact on the artifact's conclusions?
2	Known limitations declared — methodological or coverage limitations are explicitly stated	Are limitations listed (not inferred)? Does the section distinguish limitations from gaps?	Are limitations acknowledged BEFORE they are used to discount findings, or only in a footnote after conclusions are stated? A limitation that invalidates a key finding must appear near that finding, not only in the Gaps section.
3	Suggestions for further research included — the artifact proposes next steps to close identified gaps	Are 1+ follow-up research suggestions present? Are they tied to specific gaps?	Are suggestions actionable (specific query + source type) or vague ("investigate further")? Do suggestions address the gaps that most affect decision-making?

6. Cross-Validation

#	Check	Presence	Quality Depth
1	Key claims verified by 2+ sources — core conclusions are supported by multiple independent sources	Are 2+ sources cited for each key claim?	Are the sources truly independent, or do they cite each other (amplification, not validation)? Two sources from the same original study do not constitute cross-validation.
2	Discrepancies noted where sources conflict — when sources contradict each other, the conflict is documented	Are source conflicts recorded in the artifact?	Are conflicts resolved or just noted? If sources conflict, the artifact should explain which source was weighted more and why — not just acknowledge the conflict exists.
3	Single-source claims marked as unverified — any finding backed by only one source is explicitly labeled	Are single-source claims identified with an "unverified" marker or equivalent?	Are single-source claims given confidence scores below 60% as required? A single source labeled "unverified" but assigned 75% confidence is self-contradictory.

7. Actionability

#	Check	Presence	Quality Depth
1	Recommendations are concrete (not vague) — each recommendation specifies who should do what	Are recommendations present? Do they name an action (not just "consider X")?	Are recommendations actionable enough to assign to a specific role without further clarification? "Improve monitoring" is not actionable; "add P99 latency alert at 500ms threshold" is.
2	Next steps are evidence-based — proposed actions are traceable to findings in the artifact	Are next steps present and linked to specific findings?	Is each next step traceable to a specific finding and confidence score? A next step driven by a 40%-confidence finding should be labeled speculative, not presented as a directive.
3	Executive summary captures key findings — the summary conveys findings accurately without omitting critical caveats	Is an executive summary present? Does it list key findings?	Does the summary preserve confidence caveats and limitations, or does it strip them out for readability? A summary that presents 60%-confidence findings as facts is worse than no summary.

Output Format

## Knowledge Review Result

**Status:** PASS | WARN | FAIL
**Artifact:** {path}

### Checks (N/7 passed)

- [x] Template compliance
- [x] Citation audit
- [ ] Confidence accuracy — {issue}
- [ ] Source quality — {issue}
      ...

### Issues

- {specific issues found}

### Verdict

{APPROVED | REVISE | BLOCKED}

Round 2: Focused Re-Review (conditional — triggered by findings)

Convergence follows the single contract in SYNC:double-round-trip-review below: a clean Round 1 ENDS the review — there is no unconditional mandatory Round 2. Re-review is triggered by a validated-finding fix cycle (review → validate findings → fix → full re-review), not by a round number.

When Round 1 surfaces findings, run this focused re-review as part of that full re-review (do NOT rely on Round 1 memory):

1. Re-read the Round 1 verdict and checklist results
2. Re-evaluate ALL checklist items from scratch
3. Challenge Round 1 PASS items: "Is this really PASS? Did I verify citations and confidence?"
4. Focus on what Round 1 typically misses:
   • Citation accuracy (do sources actually say what's claimed?)
   • Confidence calibration (are percentages realistic?)
   • Knowledge gaps that weren't flagged
   • Template compliance shortcuts
5. Update verdict to incorporate the new findings; then re-enter the loop until a complete review pass finds zero issues.

---

[IMPORTANT] Use TaskCreate to break ALL work into small tasks BEFORE starting.

Prerequisites: MUST ATTENTION READ before executing:

OOP & DRY Enforcement: MANDATORY — flag duplicated patterns that should be extracted to a base class, generic, or helper. Classes in the same group or suffix (ex *Entity, *Dto, *Service, etc...) must inherit a common base (even if empty now — enables future shared logic and child overrides). Verify project has code linting/analyzer configured for the stack.

AI Mistake Prevention — Failure modes to avoid on every task:

Re-read files after context changes. Context compaction, resume, or long-running work can make memory stale; verify current files before acting. Verify generated content against source evidence. AI hallucinates APIs, names, claims, and document facts. Check the relevant source before documenting or referencing. Check downstream references before deleting or renaming. Removing an artifact can stale docs, generated mirrors, configs, and callers; map references first. Trace the full impact chain after edits. Changing a definition can miss derived outputs and consumers. Follow the affected chain before declaring done. Verify ALL affected outputs, not just the first. One green check is not all green checks; validate every output surface the change can affect. Assume existing values are intentional — ask WHY before changing. Before changing a constant, limit, flag, wording, or pattern, read nearby context and history. Surface ambiguity before acting — don't pick silently. Multiple valid interpretations require an explicit question or stated assumption with risk. Keep shared guidance role-relevant. Universal guidance must help every receiving skill or agent; code-specific obligations belong only in code-specific protocols.

Validated-Finding Fix + Full Re-Review Loop — Re-review is triggered by a validated finding fix cycle, not by a round number. Review purpose: review → validate findings → fix validated findings → full re-review until a complete review pass finds no issues. A clean review ENDS the loop — no further rounds required.

aka Self-Review Convergence Loop. The name is historical — there is NO 2-round cap; "double-round-trip" only means a validated-finding fix cycle forces at least one fresh re-review. It runs until a clean pass, bounded by the 5-round ceiling below.

Round cap — 5 rounds MAX (a ceiling, NEVER a target). A clean pass ENDS the loop immediately at ANY round — round 1 included; the cap never obliges you to keep spinning. Hitting round 5 with validated findings still open → STOP and escalate via AskUserQuestion with the still-open findings listed; NEVER emit a silent "good enough" PASS on cap exhaustion, and NEVER let the cap substitute for the clean-review requirement. The 3-repeated-no-progress blocker rule stays an EARLIER exit — escalate at whichever trips first.

Universal scope (any new output/judgment): any newly produced output or judgment gets ≥1 self-review; any new judgment gets ≥1 /why-review --validate-findings pass; anything flagged to re-check is re-checked ≥1 time — before that output is treated as final. This loop is the default convergence contract for ANY work-producing skill, not review skills only.

Routing invariant (author-facing): a skill that validates findings MUST route them through /why-review --validate-findings (the terminal validator) — NEVER fork an inline finding-validation. Routing through why-review is what makes the finding-survival bar and this loop apply; the verify-review-validate-coverage sensor enforces this exact route mechanically.

Round 1: Main-session review. Read target files, build understanding, note issues. Output findings + verdict (PASS / FAIL).

Decision after Round 1:

• No issues found (PASS, zero findings) → review ENDS. Do NOT spawn a fresh sub-agent for confirmation.
• Issues found (FAIL, or any non-zero findings) → run the active review skill's findings-validation gate first; for review skills the default gate is /why-review --validate-findings <report-path>. Fix only validated findings, then restart the full review protocol from the beginning with a fresh task breakdown.

Fresh full re-review after every fix cycle: Re-run the whole review protocol over the current full target. When sub-agents are part of that protocol, spawn NEW Agent calls — never reuse prior agents. Reviewers re-read ALL files from scratch with ZERO memory of prior rounds. See SYNC:fresh-context-review for the spawn mechanism and SYNC:review-protocol-injection for the canonical Agent prompt template. Each fresh full review must catch:

• Cross-cutting concerns missed in the prior round
• Interaction bugs between changed files
• Convention drift (new code vs existing patterns)
• Missing pieces that should exist but don't
• Subtle edge cases the prior round rationalized away
• Regressions introduced by the fixes themselves

Loop termination: After each full re-review, repeat the same decision: clean → END; issues → validate findings → fix → restart from the first review phase. Continue until a complete review pass finds zero issues, capped at 5 rounds. Escalate via AskUserQuestion at whichever comes first: the same validated finding repeats for 3 full invocations with no progress · a fix requires product/owner input · round 5 completes with validated findings still open. NEVER loop past 5 rounds, and NEVER convert cap exhaustion into a PASS.

Rules:

• A clean Round 1 ENDS the review — no mandatory Round 2
• NEVER fix unvalidated findings; validate first using the caller's validation gate
• Every surviving finding must additionally clear the finding-survival bar defined in why-review's Findings Validation Routine (a deliberately higher bar than the generic act-gate — "keep this finding?" is a stricter question than "act on this evidence?"); a finding below the bar is demoted or dropped, not kept
• NEVER skip the full re-review after a fix cycle (every fix invalidates the prior verdict)
• NEVER reuse a sub-agent across rounds — every iteration that uses sub-agents spawns NEW Agent calls
• Main agent READS sub-agent reports but MUST NOT filter, reinterpret, or override findings
• The 5-round cap NEVER replaces the clean-review requirement — it bounds runaway looping, it does not authorize shipping an un-clean review; a clean pass ends the loop early at any round, and cap exhaustion escalates rather than passes
• Enforce the round cap of 5 alongside the 3 repeated-no-progress blocker rule; both are escalation triggers, neither is a completion criterion
• Track recursive invocation count and repeated blockers in conversation context (session-scoped)
• Final verdict must incorporate ALL rounds executed

Report must include ## Round N Findings (Fresh Sub-Agent) for every round N≥2 that was executed.

Fresh Context Re-Review — Eliminate orchestrator confirmation bias after fixes by restarting the full review with isolated sub-agents where applicable.

Why: The main agent knows what it (or /feature-implement) just fixed and rationalizes findings accordingly. A fresh sub-agent has ZERO memory, re-reads from scratch, and catches what the main agent dismissed. Sub-agent bias is mitigated by (1) fresh context, (2) verbatim protocol injection, (3) main agent not filtering the report.

When: ONLY after a validated-finding fix cycle. A review round that finds zero issues ENDS the loop — do NOT spawn a confirmation sub-agent. A review round that finds issues triggers: validate findings → fix → full review restart from the first phase.

How:

1. Start a NEW full review invocation/task breakdown; when that protocol calls for agents, spawn NEW Agent tool calls — use code-reviewer subagent_type for code reviews, general-purpose for plan/doc/artifact reviews
2. Inject ALL required review protocols VERBATIM into the prompt — see SYNC:review-protocol-injection for the full list and template. Never reference protocols by file path; AI compliance drops behind file-read indirection (see SYNC:shared-protocol-duplication-policy)
3. Sub-agent re-reads ALL target files from scratch via its own tool calls — never pass file contents inline in the prompt
4. Sub-agent writes structured report to plans/reports/{review-type}-round{N}-{date}.md
5. Main agent reads the report, integrates findings into its own report, DOES NOT override or filter

Rules:

• SKIP fresh sub-agent when the prior full review found zero issues (no fixes = nothing new to verify)
• NEVER skip the full review restart after a fix cycle — every fix invalidates the prior verdict
• NEVER reuse a sub-agent across rounds — every fresh round spawns a NEW Agent call
• Continue until a complete full review pass has zero findings; if the same blocker repeats 3 times with no progress, escalate via AskUserQuestion
• Track iteration count and repeated blockers in conversation context (session-scoped, no persistent files)

Review Protocol Injection — Every fresh sub-agent review prompt MUST embed 11 protocol blocks VERBATIM. The template below has ALL 11 bodies already expanded inline. Copy the template wholesale into the Agent call's prompt field at runtime, replacing only the {placeholders} in Task / Round / Reference Docs / Target Files / Output sections with context-specific values. Do NOT touch the embedded protocol sections.

Why inline expansion: Placeholder markers would force file-read indirection at runtime. AI compliance drops significantly behind indirection (see SYNC:shared-protocol-duplication-policy). Therefore the template carries all 11 protocol bodies pre-embedded.

Subagent Type Selection

• code-reviewer — for code reviews (reviewing source files, git diffs, implementation)
• general-purpose — for plan / doc / artifact reviews (reviewing markdown plans, docs, specs)

Canonical Agent Call Template (Copy Verbatim)

Agent({
  description: "Fresh Round {N} review",
  subagent_type: "code-reviewer",
  prompt: `
## Task
{review-specific task — e.g., "Review all uncommitted changes for code quality" | "Review plan files under {plan-dir}" | "Review integration tests in {path}"}

## Round
Round {N}. You have ZERO memory of prior rounds. Re-read all target files from scratch via your own tool calls. Do NOT trust anything from the main agent beyond this prompt.

## Protocols (follow VERBATIM — these are non-negotiable)

### Spec ↔ Tests ↔ Code Triangulation
DO THIS FIRST — before any per-protocol check below. The review target is the WHOLE PACKAGE, not the diff alone: load the behavior's spec (§3 ACs / §4 BRs / §8 TCs), its tests, and the changed code TOGETHER, and reason about their mutual consistency BEFORE judging any one in isolation.
1. Locate all three faces: the Feature Spec section(s) governing the changed behavior, the tests that guard it, and the production code that implements it. A missing face is itself a finding (SPEC-GAP / TEST-GAP / DEAD-SPEC).
2. Triangulate pairwise — every disagreement is a finding; classify which face is wrong:
   - code vs spec: behavior the code does that no §3/§4/§8 rule describes → CODE-EXTRA or SPEC-STALE; a [HARD] §4 rule or §5 invariant with no enforcing code path → CODE-WRONG.
   - tests vs spec: a §8 TC with no test, or a test asserting behavior no TC/rule names → TEST-GAP or SPEC-SILENT.
   - tests vs code: a changed code path with no covering test → TEST-GAP; a test that still passes against a deliberately broken invariant → WEAK-TEST (apply the mutation thinking in Bug Detection).
3. Hidden-rule capture: any invariant the code enforces but the spec never states (SPEC-SILENT) MUST be surfaced as a finding to add into §3/§4/§8 AND guarded with a test — the enrichment loop, never a silent pass.
4. Only after the three faces agree — or every disagreement is logged as a finding — proceed to the per-protocol checks below; when enrichment adds spec/test content, re-review the package against the enriched spec.
NEVER mark review PASS while any spec/test/code face disagrees without a logged finding. The diff is the entry point; the package is the unit of judgment.

### Evidence-Based Reasoning
Speculation is FORBIDDEN. Every claim needs proof.
1. Cite file:line, grep results, or framework docs for EVERY claim
2. Declare confidence: >80% act freely, 60-80% verify first, <60% DO NOT recommend
3. Cross-service validation required for architectural changes
4. "I don't have enough evidence" is valid and expected output
BLOCKED until: Evidence file path (file:line) provided; Grep search performed; 3+ similar patterns found; Confidence level stated.
Forbidden without proof: "obviously", "I think", "should be", "probably", "this is because".
If incomplete → output: "Insufficient evidence. Verified: [...]. Not verified: [...]."

### Bug Detection
MUST check categories 1-4 for EVERY review. Never skip.
1. Null Safety: Can params/returns be null? Are they guarded? Optional chaining gaps? .find() returns checked?
2. Boundary Conditions: Off-by-one (< vs <=)? Empty collections handled? Zero/negative values? Max limits?
3. Error Handling: Try-catch scope correct? Silent swallowed exceptions? Error types specific? Cleanup in finally?
4. Resource Management: Connections/streams closed? Subscriptions unsubscribed on destroy? Timers cleared? Memory bounded?
5. Concurrency (if async): Missing await? Race conditions on shared state? Stale closures? Retry storms?
6. Stack-Specific: Check the configured language/runtime pitfalls and framework-specific failure modes discovered from local code.
Classify: CRITICAL (crash/corrupt) → FAIL | HIGH (incorrect behavior) → FAIL | MEDIUM (edge case) → WARN | LOW (defensive) → INFO.

### Design Patterns Quality
Priority checks for every code change:
1. DRY via OOP: Same-suffix classes (*Entity, *Dto, *Service) MUST share base class. 3+ similar patterns → extract to shared abstraction.
2. Right Responsibility: Logic in LOWEST layer (Entity > Domain Service > Application Service > Controller). Never business logic in controllers.
3. SOLID: Single responsibility (one reason to change). Open-closed (extend, don't modify). Liskov (subtypes substitutable). Interface segregation (small interfaces). Dependency inversion (depend on abstractions).
4. After extraction/move/rename: Grep ENTIRE scope for dangling references. Zero tolerance.
5. YAGNI gate: NEVER recommend patterns unless 3+ occurrences exist. Don't extract for hypothetical future use.
Anti-patterns to flag: God Object, Copy-Paste inheritance, Circular Dependency, Leaky Abstraction.

### Logic & Intention Review
Verify WHAT code does matches WHY it was changed.
1. Change Intention Check: Every changed file MUST serve the stated purpose. Flag unrelated changes as scope creep.
2. Happy Path Trace: Walk through one complete success scenario through changed code.
3. Error Path Trace: Walk through one failure/edge case scenario through changed code.
4. Acceptance Mapping: If plan context available, map every acceptance criterion to a code change.
5. Tests Verify Intent: For test/spec changes, verify tests name the protected business rule or invariant and would fail if that intent breaks.
6. Migration Test Exclusion: Do not write tests for migration code. Schema/data migrations are one-time execution paths, not core application logic.
NEVER mark review PASS without completing both traces (happy + error path).

### Test Spec Verification
Map changed code to test specifications.
1. Identify the project's test/spec format from existing docs, test-case files, BDD feature files, or spec folders.
2. Every changed code path MUST map to a corresponding test case/spec (or flag as "needs test case").
3. New functions/endpoints/handlers → flag for test spec creation.
4. Migration files are excluded from test/spec creation; schema/data migrations are one-time execution paths, not core application logic.
5. If spec evidence fields exist, verify they point to actual code (file:line, not stale references).
6. Verify each meaningful test case names the business intent/invariant; flag behavior-only cases that only mirror implementation details.
7. Auth/data changes → verify corresponding authorization and data-state test cases exist.
8. If no specs exist for a changed path → log the gap and recommend the project's test-spec workflow.
NEVER skip test mapping. Untested code paths are the #1 source of production bugs.

### Behavioral Delta Matrix
MANDATORY for any bugfix review. Produce input-state × pre-fix × post-fix × delta table BEFORE writing verdict.
- Minimum 3 rows; include at least one row OUTSIDE the original bug report.
- Any "REGRESSION" delta → review returns FAIL until a preservation test is added.
- Narrative descriptions do NOT substitute for the matrix.
Example rows (external-record sync fix):
| Input                 | Pre-fix | Post-fix                  | Delta      |
| --------------------- | ------- | ------------------------- | ---------- |
| Record exists (valid) | Reused  | Always recreated → orphan | REGRESSION |
| Record missing (404)  | Error   | Recreated                 | Fixed      |

### Fix-Layer Accountability
NEVER fix at the crash site. Trace the full flow, fix at the owning layer. The crash site is a SYMPTOM, not the cause.
MANDATORY before ANY fix:
1. Trace full data flow — Map the complete path from data origin to crash site across ALL layers (storage → backend → API → frontend → UI). Identify where bad state ENTERS, not where it CRASHES.
2. Identify the invariant owner — Which layer's contract guarantees this value is valid? Fix at the LOWEST layer that owns the invariant, not the highest layer that consumes it.
3. One fix, maximum protection — If fix requires touching 3+ files with defensive checks, you are at the wrong layer — go lower.
4. Verify no bypass paths — Confirm all data flows through the fix point. Check for direct construction skipping factories, clone/spread without re-validation, raw data not wrapped in domain models, mutations outside the model layer.
BLOCKED until: Full data flow traced (origin → crash); Invariant owner identified with file:line evidence; All access sites audited (grep count); Fix layer justified (lowest layer that protects most consumers).
Anti-patterns (REJECT): "Fix it where it crashes" (crash site ≠ cause site, trace upstream); "Add defensive checks at every consumer" (scattered defense = wrong layer); "Both fix is safer" (pick ONE authoritative layer).

### Rationalization Prevention
AI skips steps via these evasions. Recognize and reject:
- "Too simple for a plan" → Simple + wrong assumptions = wasted time. Plan anyway.
- "I'll test after" → RED before GREEN. Write/verify test first.
- "Already searched" → Show grep evidence with file:line. No proof = no search.
- "Just do it" → Still need TaskCreate. Skip depth, never skip tracking.
- "Just a small fix" → Small fix in wrong location cascades. Verify file:line first.
- "Code is self-explanatory" → Future readers need evidence trail. Document anyway.
- "Combine steps to save time" → Combined steps dilute focus. Each step has distinct purpose.

### Graph-Assisted Investigation
MANDATORY when .code-graph/graph.db exists.
HARD-GATE: MUST run at least ONE graph command on key files before concluding any investigation.
Pattern: Grep finds files → trace --direction both reveals full system flow → Grep verifies details.
- Investigation/Scout: trace --direction both on 2-3 entry files
- Fix/Debug: callers_of on buggy function + tests_for
- Feature/Enhancement: connections on files to be modified
- Code Review: tests_for on changed functions
- Blast Radius: trace --direction downstream
CLI: python .claude/scripts/code_graph {command} --json. Use --node-mode file first (10-30x less noise), then --node-mode function for detail.

### Understand Code First
HARD-GATE: Do NOT write, plan, or fix until you READ existing code.
1. Search 3+ similar patterns (grep/glob) — cite file:line evidence.
2. Read existing files in target area — understand structure, base classes, conventions.
3. Run python .claude/scripts/code_graph trace <file> --direction both --json when .code-graph/graph.db exists.
4. Map dependencies via connections or callers_of — know what depends on your target.
5. Write investigation to .ai/workspace/analysis/ for non-trivial tasks (3+ files).
6. Re-read analysis file before implementing — never work from memory alone.
7. NEVER invent new patterns when existing ones work — match exactly or document deviation.
BLOCKED until: Read target files; Grep 3+ patterns; Graph trace (if graph.db exists); Assumptions verified with evidence.

## Reference Docs (READ before reviewing)
- `.claude/docs/development-rules.md` — canonical development rules, code-quality guidelines, and pre-commit checklist
- docs/project-reference/code-review-rules.md
- {skill-specific reference docs — e.g., integration-test-reference.md for integration-test-review; backend-patterns-reference.md for backend reviews; frontend-patterns-reference.md for frontend reviews}

## Target Files
{explicit file list OR "run git diff to see uncommitted changes" OR "read all files under {plan-dir}"}

## Output
Write a structured report to plans/reports/{review-type}-round{N}-{date}.md with sections:
- Status: PASS | FAIL
- Issue Count: {number}
- Critical Issues (with file:line evidence)
- High Priority Issues (with file:line evidence)
- Medium / Low Issues
- Cross-cutting findings

Return the report path and status to the main agent.
Every finding MUST have file:line evidence. Speculation is forbidden.
`
})

Rules

• DO copy the template wholesale — including all 11 embedded protocol sections
• DO replace only the {placeholders} in Task / Round / Reference Docs / Target Files / Output sections with context-specific content
• DO choose code-reviewer subagent_type for code reviews and general-purpose for plan / doc / artifact reviews
• DO NOT paraphrase, summarize, or skip any protocol section
• DO NOT pass file contents inline — the sub-agent reads via its own tool calls so it has a fresh context
• DO NOT reference protocols by file path or tag name — the bodies are already embedded above
• DO NOT introduce placeholder markers for the protocols — they must stay literally expanded

Nested Task Expansion Contract — For workflow-step invocation, the [Workflow] ... row is only a parent container; the child skill still creates visible phase tasks.

1. Call TaskList first. If a matching active parent workflow row exists, set nested=true and record parentTaskId; otherwise run standalone.
2. Create one task per declared phase before phase work. When nested, prefix subjects [N.M] $skill-name — phase.
3. When nested, link the parent with TaskUpdate(parentTaskId, addBlockedBy: [childIds]).
4. Orchestrators must pre-expand a child skill's phase list and link the workflow row before invoking that child skill or sub-agent.
5. Mark exactly one child in_progress before work and completed immediately after evidence is written.
6. Complete the parent only after all child tasks are completed or explicitly cancelled with reason.

Blocked until: TaskList done, child phases created, parent linked when nested, first child marked in_progress.

Project Reference Docs Gate — Run after task-tracking bootstrap and before target/source file reads, grep, edits, or analysis. Project docs override generic framework assumptions.

1. Identify scope: file types, domain area, and operation.
2. Read docs/project-config.json first — the project's machine-readable map. It is the single source of truth for THIS repo (modules/paths, framework + search keywords, test/E2E/integration run-commands, design system, architecture rules, workflow patterns); ground exact paths, run-commands, and conventions on it before investigating, planning, or coding — never assume framework defaults (CLAUDE.md + reference docs are derived from it). If it — or the docs index, lessons.md, CLAUDE.md, AGENTS.md, or any required reference doc — is missing or stale, auto-run /project-init or the narrow route (/project-config, /docs-init, /scan-all, /scan --target=<key>, /claude-md-init) first; if Codex mirrors or AGENTS.md are stale, ask the user to run /sync-codex (never auto-run it).
3. Required docs by trigger: always docs/project-reference/lessons.md; doc lookup docs-index-reference.md; review code-review-rules.md; backend/CQRS/API backend-patterns-reference.md; domain/entity domain-entities-reference.md; frontend/UI frontend-patterns-reference.md; styles/design scss-styling-guide.md + design-system/design-system-canonical.md; integration tests integration-test-reference.md; E2E e2e-test-reference.md; feature docs/specs feature-spec-reference.md + spec-system-reference.md + spec-principles.md; behavior/public-contract/spec-test-code sync workflow-spec-test-code-cycle-reference.md; derived spec index/ERD/reimplementation guides spec-system-reference.md + source Feature Specs under docs/specs/; architecture/new area project-structure-reference.md.
4. Read every required doc, then before target work state: Reference docs read: ... | Not applicable: ....

Ready when: scope evaluated, docs/project-config.json consulted, required docs checked/read or setup route completed, lessons.md confirmed, citation emitted.

Task Tracking & External Report Persistence — Bootstrap this before execution; then run project-reference doc prefetch before target/source work.

1. Create a small task breakdown before target file reads, grep, edits, or analysis. On context loss, inspect the current task list first.
2. Mark one task in_progress before work and completed immediately after evidence; never batch transitions.
3. For plan/review work, create plans/reports/{skill}-{YYMMDD}-{HHmm}-{slug}.md before first finding.
4. Append findings after each file/section/decision and synthesize from the report file at the end.
5. Final output cites Full report: plans/reports/{filename}.

Blocked until: task breakdown exists, report path declared for plan/review work, first finding persisted before the next finding.

Critical Thinking Mindset — Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence >80% to act. Anti-hallucination: Never present guess as fact — cite sources for every claim, admit uncertainty freely, self-check output for errors, cross-reference independently, stay skeptical of own confidence — certainty without evidence root of all hallucination.

Web Research — Structured web search for evidence gathering.

1. Form 3-5 specific search queries (not generic questions)
2. Use WebSearch for each query, collect top 3-5 sources
3. Validate source credibility (official docs > blogs > forums)
4. Cross-validate claims across 2+ sources before citing
5. Write findings to research report with source URLs

NEVER cite a single source as authoritative. Always cross-validate.

Severity Rubric — Classify every finding by consequence, not by how easy it is to fix. One scale across all reviews so a "High" means the same thing everywhere.

Severity	Action	Definition
CRITICAL	Block merge	Silent runtime failure, data corruption, validation bypass, security hole
HIGH	Must fix	Incorrect behavior, invariant gap, architectural violation
MEDIUM	Should fix	Design debt, maintainability, likely future bug
LOW	Nice to fix	Convention, documentation, minor clarity

Score-based skills map their numeric scale onto these tiers — do not invent a parallel vocabulary:

• 0-2 criterion scoring (e.g. production-readiness-review): 0 = CRITICAL/HIGH (criterion unmet, blocks production readiness), 1 = MEDIUM (partial, should fix), 2 = pass (no finding).
• Two-axis scoring (e.g. performance-review, impact × likelihood): map the resulting cell to the nearest tier — high-impact + high-likelihood → CRITICAL/HIGH; low-impact OR low-likelihood → MEDIUM/LOW.

A finding's tier drives the gate: CRITICAL/HIGH must be resolved or explicitly accepted by the owner before PASS; MEDIUM/LOW may ship with a tracked follow-up.

Goal Contract Satisfaction Loop — Persist the user goal in an external file, execute against it, and loop review/fix until every saved required criterion passes or a blocker escalates. Bounded closed loop — NEVER open-ended autonomous exploration.

1. Resolve the active goal (in order): active plan goal.md → plans/goals/{YYMMDD-HHmm}-{slug}/goal.md → create a new Goal Contract from the current user request (template: .claude/templates/goal-contract-template.md).
2. Required sections: Original Request, Purpose, Success Criteria (checkboxes; mark required vs optional), Constraints, Evidence Required, Iteration Log, Goal Satisfaction matrix.
3. Before work: read the active goal and map planned work to saved success criteria — execution serves the saved criteria, never chat memory alone.
4. After execution/verification: append an Iteration Log entry — result, evidence references (file:line, command output, report path), remaining gaps.
5. Review gate: emit a Goal Satisfaction matrix — | Success Criterion | Evidence | Status | with PASS/FAIL/BLOCKED. Overall PASS requires every required criterion PASS.
6. Loop rule (retry): required criterion FAIL → validate the gap is real → fix → re-review only the affected criteria. Stop cleanly when all required criteria PASS.
7. Escalation rule (stop): two consecutive iterations with no criterion progressing, or a blocker needing user input → mark the criterion BLOCKED with a user-facing reason and escalate. NEVER loop indefinitely.
8. Skip rule: tiny conversational tasks may skip the goal file ONLY with a recorded one-line reason. User-accepted gate skips are recorded in the goal file with reason and scope.
9. Security: NEVER store secrets, tokens, credentials, or private customer data in goal files — store evidence references and redact sensitive values.

Blocked until: active goal resolved (or skip reason recorded) · saved success criteria read before edits · iteration evidence appended after execution · Goal Satisfaction matrix emitted before any PASS verdict.

Trade-Off Interrogation Gate — ALWAYS ask these THREE questions before ANY verdict, score, finding, or recommendation — about the thing under review AND about every recommendation YOU make. — why: naming a benefit without its price is an endorsement, not a review; the costliest trade-offs are the ones nobody wrote down.

1. Is there any trade-off? Name what it SACRIFICES. "None" / "pure win" is an unfinished analysis, NOT an answer — to claim none, state which dimensions you checked and why each is unaffected: future change cost · complexity · performance/latency · memory/cost · coupling · reversibility · migration burden · operational load · blast radius · security posture · testability · team skill/ramp · delivery time · UX.
2. Is it worth it? Weigh gain against sacrifice EXPLICITLY — what is gained (with a metric) · what it costs · WHO pays · WHEN it comes due — then emit WORTH IT / NOT WORTH IT / UNCLEAR. "Better" with no metric and no cost FAILS this question. NOT WORTH IT → withdraw or replace the recommendation, never keep it as-is.
3. Is the trade-off material enough to CONFIRM WITH THE USER? A material trade-off is the user's call, never yours. MATERIAL when ANY holds: irreversible / one-way door (data migration, public contract, storage format, vendor lock-in) · cost shifted onto someone else (another team, ops/on-call, future maintainer, end user) · one quality attribute traded for another (correctness↔speed, security↔convenience, latency↔cost, simplicity↔flexibility) · a boundary crossed (client↔server tier, service contract, event contract, shared library) · a high-consequence path (auth, money, data integrity, breaking change, High/Medium residual risk) · the worth-it verdict is UNCLEAR.

MATERIAL → STOP and confirm via AskUserQuestion BEFORE the verdict stands — state the trade-off, both options, what each sacrifices, and your recommendation. NOT material → record it inline with a one-line justification and proceed.

Non-asking execution contexts — ESCALATE BY HANDOFF, never by silence. AskUserQuestion reaches only the main interactive agent: a sub-agent cannot ask the user, and a terminal/verdict-only mode asks nothing by design. When you are running in such a context, the obligation is redirected, never waived — do ALL of: (a) complete questions 1 and 2 normally; (b) decide materiality and record it in the Trade-Off Assessment row with confirmed? = NO — cannot ask from this context; (c) name the unconfirmed MATERIAL trade-off explicitly in your returned summary/verdict so the CALLER (or parent orchestrator) escalates it via AskUserQuestion on your behalf — a material trade-off mentioned only inside a report file on disk is NOT a handoff; (d) do not emit an unqualified PASS — mark the verdict as carrying an unconfirmed material trade-off, so the caller's gate stays closed until the user answers. The caller inherits the escalation duty the moment it reads your return.

This carve-out is about reachability, not convenience: it applies ONLY where the tool genuinely cannot reach the user (spawned sub-agent, terminal validate/verdict-only mode, non-interactive/headless run). It is NEVER a licence to skip the question, to self-approve a one-way door, or to downgrade materiality because asking is inconvenient — if you CAN ask, you MUST ask.

Emit a Trade-Off Assessment row per reviewed decision and per recommendation: | decision | sacrifices | gain (metric) | who pays, when | WORTH IT/NOT/UNCLEAR | material? | confirmed? |.

BLOCKED until: trade-off named (or dimensions-checked justification given) · worth-it verdict emitted · materiality decided · every MATERIAL trade-off either confirmed with the user OR — in a non-asking context — handed off in the returned verdict for the caller to confirm. A MATERIAL trade-off that is neither confirmed nor handed off can NEVER be PASS, and NEVER gets buried as a Low-severity note.

NEVER answer "no trade-off" without checking · decide a material trade-off silently on the user's behalf · let convergence/delivery pressure authorize walking through a one-way door · bundle several material trade-offs into one vague "proceed?".

• MANDATORY IMPORTANT MUST ATTENTION execute the review loop (aka Self-Review Convergence Loop): review → validate findings → fix validated findings → full re-review. A complete review pass with zero findings ENDS the review. Any newly produced output/judgment gets ≥1 self-review; any new judgment gets ≥1 /why-review --validate-findings pass before it is treated as final.
• MANDATORY enforce the round cap of 5 — a ceiling, NEVER a target: a clean pass ends the loop immediately at any round (round 1 included), and round 5 completing with validated findings still open → STOP & escalate via AskUserQuestion, never a silent PASS. The 3-repeated-no-progress blocker rule is an earlier exit — escalate at whichever trips first. NEVER loop open-ended.

IMPORTANT MUST ATTENTION cite 2+ independent sources per claim. NEVER fabricate — "No evidence found" is valid output.

MUST ATTENTION apply critical + sequential thinking — every claim needs appropriate traced evidence (file:line for repo/code claims; source URL or artifact section for research, product, content, and docs claims); confidence >80% to act, <60% DO NOT recommend. Anti-hallucination: never present guess as fact, admit uncertainty freely, cross-reference independently, stay skeptical of own confidence.

MUST ATTENTION apply AI mistake prevention — verify generated content against evidence, trace downstream references before deleting or renaming, verify all affected outputs, re-read files after context loss, and surface ambiguity before acting.

• MANDATORY Bootstrap task tracking before target work; transition one task at a time.
• MANDATORY Persist plan/review findings to plans/reports/ incrementally and synthesize from disk.

• MANDATORY Before investigating, planning, or coding, read docs/project-config.json (the project map: modules/paths, run-commands, conventions, architecture/workflow rules) + the required project-reference docs, and cite Reference docs read: ....
• MANDATORY Always include lessons.md; project config + conventions override generic framework defaults.
• MANDATORY If project config, root instruction files, or any required reference doc is missing or stale, auto-run /project-init or the narrow lower-level route before ordinary project-specific work.

• MANDATORY Parent workflow rows do not replace child phase tracking; expand phases and link the parent when nested.
• MANDATORY Orchestrators pre-expand child skill phases before invocation; use [N.M] $skill-name — phase prefixes and one-in_progress discipline.

• MANDATORY Classify findings Critical/High/Medium/Low by consequence; Critical/High block PASS until fixed or owner-accepted.
• MANDATORY Score-based skills (sre 0-2, perf two-axis) map onto the same four tiers — no parallel severity vocabulary.

• MANDATORY Resolve the active Goal Contract BEFORE work (active plan goal.md → plans/goals/{YYMMDD-HHmm}-{slug}/goal.md → create from current request) and read saved success criteria before editing.
• MANDATORY Append iteration evidence after execution; emit a Goal Satisfaction matrix (PASS/FAIL/BLOCKED) before reporting PASS; loop on validated FAIL; escalate repeated no-progress or blockers. NEVER store secrets in goal files.

• MANDATORY MUST ATTENTION ALWAYS ASK THE 3 TRADE-OFF QUESTIONS — on the thing under review AND on every recommendation you make: (1) is there any trade-off? name what it SACRIFICES (change cost · complexity · perf · coupling · reversibility · migration · ops load · blast radius · security · testability · delivery time · UX) — "none"/"pure win" is an unfinished analysis, so state the dimensions checked; (2) is it worth it? gain (with a metric) vs cost, WHO pays, WHEN → emit WORTH IT / NOT WORTH IT / UNCLEAR; NOT WORTH IT → withdraw or replace it; (3) is it material enough to confirm with the user? irreversible/one-way door · cost shifted onto another team/ops/maintainer/user · one quality attribute traded for another · a tier/service/event/library boundary crossed · auth/money/data-integrity/breaking-change/High-or-Medium-risk path · verdict UNCLEAR → STOP and confirm via AskUserQuestion BEFORE the verdict.
• MANDATORY A MATERIAL trade-off with no user confirmation can NEVER be PASS; NEVER bury one as a Low-severity note, NEVER decide it silently, and NEVER let delivery or convergence pressure authorize a one-way door. — why: an un-walked-back one-way door is the user's call to make, not the reviewer's.
• MANDATORY — non-asking contexts escalate BY HANDOFF, never by silence. AskUserQuestion reaches only the main interactive agent: a sub-agent cannot ask the user, and a terminal/verdict-only mode asks nothing by design. There the duty is REDIRECTED, not waived — still name the trade-off, still decide materiality, record confirmed? = NO — cannot ask from this context, state the unconfirmed MATERIAL trade-off in your RETURNED verdict/summary so the CALLER escalates it (a note only in an on-disk report is not a handoff), and never emit an unqualified PASS. Applies ONLY where the user is genuinely unreachable (spawned sub-agent, terminal validate mode, headless run) — if you CAN ask, you MUST ask.

Closing Reminders

IMPORTANT MUST ATTENTION Goal: Ensure knowledge artifacts are evidence-backed, complete, protocol-compliant, and safe to use for decisions — review for quality, completeness, citation accuracy, and template compliance.

Protocols in force (concise digest of the SYNC/shared blocks this skill carries):

• AI Mistake Prevention: verify generated content against evidence, trace downstream references, verify all affected outputs, re-read after context loss, surface ambiguity.
• Double Round-Trip Review: review → validate → fix → full re-review; clean pass ends the loop.
• Fresh Context Review: after a fix, re-review with zero-memory fresh sub-agents.
• Review Protocol Injection: MUST ATTENTION embed all 11 protocol bodies VERBATIM into each fresh sub-agent prompt.
• Nested Task Creation: expand child phase tasks and link the parent when nested.
• Project Reference Docs Guide: ALWAYS read required project docs (lessons.md) before target review.
• Task Tracking & External Report: bootstrap tasks; persist plan/review findings to plans/reports/ incrementally.
• Critical Thinking: apply critical + sequential thinking; every claim needs traced proof, >80% to act.
• Web Research: cite 2+ independent sources per claim; NEVER fabricate.
• Severity Rubric: classify findings Critical/High/Medium/Low by consequence; Critical/High block PASS.

IMPORTANT MUST ATTENTION default SKEPTIC, not VALIDATOR — run the full Anti-Bias Gate before ANY verdict: 1+ contradicting source per major claim, stress-test every score ≥80%, state the strongest alternative conclusion, check supporting-vs-contradicting source ratio, run a pre-mortem, argue the opposite verdict in 2+ sentences — why: AI gravitates to confirming sources, so an ungated review ships confirmation-biased verdicts as fact. IMPORTANT MUST ATTENTION calibrate confidence to evidence — a single source ≠ 80%; scores >80% need 2+ independent sources with contradicting evidence addressed; single-source claims marked unverified MUST be <60%; findings <60% flagged prominently — why: an inflated confidence score is read downstream as a fact and drives bad decisions. IMPORTANT MUST ATTENTION main steps in order — read artifact → 7-checklist audit (presence AND quality depth) → adversarial Anti-Bias Gate → emit PASS/WARN/FAIL per-check + verdict (APPROVED/REVISE/BLOCKED) → conditional Round 2 re-review — why: AI forgets the skill's own pipeline and skips the audit or the gate. IMPORTANT MUST ATTENTION verify presence AND quality depth across all 7 checklists (template compliance, citation audit, confidence accuracy, source quality, knowledge gaps, cross-validation, actionability) — a section that exists but is filler/placeholder has negative value — why: "section present" ≠ "section sound"; false confidence is worse than an honest gap. IMPORTANT MUST ATTENTION READ-ONLY — review and report, NEVER modify the audited artifact; emit fixes as findings for the author — why: a reviewer that edits the artifact destroys the independent second opinion the review exists to provide.

• break work into small todo tasks using TaskCreate BEFORE starting; add a final review todo to verify work quality
• cite evidence for every claim — for a knowledge artifact that is the supporting source citation [N] / source-table row (use file:line only for the rare code-linked claim); confidence >80% to act, <60% DO NOT recommend
• read required project-reference docs (always lessons.md) before the target review; classify findings Critical/High/Medium/Low by consequence (severity rubric) — Critical/High block PASS until fixed or owner-accepted
• verify every factual claim has an inline citation, every source in the table is referenced, no orphan citations, all 5 source fields present with accurate Tier — why: citation presence ≠ citation correctness; the cited source must actually support the specific claim
• execute the review loop: review → validate findings → fix validated findings → full re-review; a complete review pass with zero findings ENDS the review — NEVER fix unvalidated findings, NEVER skip the full re-review after a fix cycle — why: every fix invalidates the prior verdict

[TASK-PLANNING] Before acting, analyze task scope and systematically break it into small todo tasks and sub-tasks using TaskCreate.

---

Closing reminder — Easy to Change is the success metric. Every finding, test, refactor, and abstraction must answer one question: does this make the next change cheaper or more expensive? If it doesn't reduce future change cost, reject it. Coupling, hidden state, duplicated knowledge, and unclear intent are the real enemies — call them out by name.

Anti-Rationalization:

Evasion	Rebuttal
"Sources are cited"	Presence ≠ quality — verify each source actually supports the SPECIFIC claim, not just sits in the table.
"Confidence scores look reasonable"	Name what would LOWER each score ≥80%. No answer = inflated. A single source is not 80%.
"Comprehensive coverage"	State the strongest alternative conclusion and the perspective MISSING — absence of counterevidence ≠ consensus.
"Recommendations are actionable"	On what evidence, at what confidence? A directive from a 40%-confidence finding must be labeled speculative.
"Clean enough, skip the Anti-Bias Gate"	The gate is MANDATORY before any verdict — skipping it ships confirmation-biased approval as fact.
"I'll just fix the artifact while here"	READ-ONLY — emit findings; editing the artifact collapses the independent review into the authoring it audits.

IMPORTANT MUST ATTENTION Goal (recency anchor): ship only evidence-backed, complete, protocol-compliant knowledge artifacts — verdict via the Anti-Bias Gate (SKEPTIC default), confidence calibrated to source count/quality, READ-ONLY, severity-tiered findings, re-review until zero issues.