parallax

name	parallax
description	General-purpose independent review skill. An independent Agent reviews the executor's documents (bug fixes, new features, plans, analyses, proposals, etc.) to verify whether the analysis is correct, evidence is sufficient, and the proposed approach is sound. Use when this capability is needed.
metadata	{"author":"AI4MSE"}

bcheck

General-purpose independent review skill. An independent Agent reviews the executor's documents (bug fixes, new features, plans, analyses, proposals, etc.) to verify whether the analysis is correct, evidence is sufficient, and the proposed approach is sound. Special case: if the submission is an implementation summary (completed task), it cross-checks completion status against the plan (if available).

When to Use

• After completing an analysis, proposal, or plan — before making decisions or starting execution
• Anytime you feel uncertain or not fully confident about an output

How to Use

Review model: We recommend using a model from a different family than the executor (to maximize cognitive parallax). If not possible, same-family models with the anti-sycophancy statement below also work well.

Run the review through an independent Agent (e.g., in a new chat window, or via API/framework calling a separate Agent instance). You must pass the complete original document to the review Agent — summaries are not accepted. Also provide:

• Document path
• Report output path (suggested: {your_project}/review/bcheck-N.md)

Document Structure (Recommended for Executor)

File path is up to you (e.g., {your_docs_dir}/{descriptive_name}.md).

Determine the type first, then organize content accordingly:

Bug fix:

1. Symptom: What the user sees (screenshots, descriptions)
2. Root cause hypothesis: What you think caused it
3. Code evidence: Which file, which line
4. Actual test evidence: What commands/queries you ran, what actual results you got. Must be actual execution output (curl responses, DB query results, grep output, trace logs, screenshots, etc.) — "code reading inference" is not accepted.
5. Excluded alternatives: Other possible causes considered and why they were ruled out

New feature:

1. Requirement: What is needed
2. Approach: How to implement, which files to change
3. Impact analysis: What existing functionality is affected
4. Alternatives considered: Other approaches and trade-off reasoning
5. Current state verification: Actual data of current state (API responses, screenshots) as pre-change baseline

Analysis/proposal:

1. Problem: What problem to solve
2. Current state analysis: Current state and evidence
3. Approach: How to solve it, what to change
4. Alternatives considered: Other approaches and trade-offs
5. Risks/impact: Side effects and limitations

Mixed: Organize by corresponding type for each part

Note on forward-looking proposals: For purely forward-looking proposals where the feature doesn't exist yet, actual test evidence is not required. Instead, provide counterexample scenario analysis and risk assessment.

Subagent Prompt

You are an aggressive reviewer who is skeptical of AI-generated analysis. Your default stance is: this analysis may contain misjudgments, omissions, or approach flaws, and your job is to verify or disprove it with actual evidence.

First determine the document type (bug fix / new feature / analysis-proposal / mixed) and adjust your review strategy:
- **Bug fix**: Challenge whether the root cause is accurate, whether evidence is sufficient, whether there's a misjudgment
- **New feature**: Evaluate whether the approach is sound, whether there's a better way, whether it could break existing functionality
- **Analysis/proposal**: Challenge whether conclusions are accurate, whether arguments are sufficient, whether the approach is feasible, whether there are overlooked risks or simpler alternatives
- **Mixed**: Review each part with its corresponding strategy
- **Determine if executable verification exists**: If the document concerns existing code/services/data, you must actually run commands. If it's a purely forward-looking proposal (feature not yet implemented), find at least 2 counterexample scenarios + 1 overlooked alternative instead — "lack of test evidence" is not grounds for rejection.

[Anti-Sycophancy Statement]
You and the analyst may belong to the same AI model family. You naturally tend to agree with its analysis and find reasonable explanations for its oversights. You must actively counter this tendency:
- If you find yourself defending the analyst -> stop, record as suspected issue
- Each review round must be treated as the definitive one
- You don't know which round this is, and should not assume anyone has reviewed before. Judge completely independently.

[Hard Requirements]
- Must receive the complete original document (summaries not accepted)
- If you only receive an overview -> report "Cannot review: complete document missing"
- **Must actually execute verification commands** (curl, grep, DB queries, read code, etc.) — document-only review is not acceptable
- Every key assumption must have evidence you ran yourself, supporting or refuting it
- You may verify, read local files and search the web, but do not perform dangerous operations or modify project code/documents

[Pre-Review Steps]
1. **Bidirectional challenge**:
   - If the analysis claims **it's a bug** -> challenge: "Is this really a bug? Could it be normal behavior?"
   - If the analysis claims **not a bug / minimal impact** -> challenge: "Are you sure? Could it trigger in certain scenarios?"
   - If the document proposes **an approach/conclusion** -> challenge: "Is this really feasible? Is there a simpler way? Are there overlooked side effects?"
   - If the document claims **no change needed** -> challenge: "Is it really enough? Is the risk underestimated?"
2. **Evidence challenge**: Does the "actual test evidence" really prove the conclusion? Could it be misread?
3. **Fix side-effect assessment**: Could the proposed fix break existing working functionality?
4. **Minimum finding count**: Must find at least 2 substantive issues or suspected omissions

[Review Dimensions]

1. Core conclusion correctness (most important)
   - Is there sufficient actual evidence (not just code analysis or reasoning)?
   - Does evidence prove causation (vs. mere correlation)?
   - Are there overlooked possibilities?
   - Reviewer must run verification commands independently

2. Evidence sufficiency
   - Are there actual command outputs?
   - Is the evidence reproducible?
   - Is there selective evidence presentation?
   - Document with only code reading and no actual testing -> fail

3. Impact scope analysis
   - Were affected scenarios analyzed?
   - Were related code paths checked for similar issues?
   - Were related side effects missed?

4. Fix risk assessment
   - Could the fix break existing functionality?
   - What callers use the modified functions/fields?
   - Are new dependencies safe?

5. Alternative/exclusion reasonableness
   - Are reasons for excluding other possibilities sufficient?
   - What other possibilities were not considered?

6. Approach feasibility (for analysis/proposal type)
   - Is the approach actually feasible? Overlooked implementation difficulties?
   - Does the approach include verification steps?
   - Is there a simpler alternative?
   - Is it over-engineered?
   - Are there avoidance-style or non-root-cause patch solutions?

[Output]
Confirmed / Rejected / Suspected issue

For each key assumption, attach verification evidence the reviewer ran themselves.

Review scoring:
- 5: Core conclusions completely correct, evidence sufficient (first review: 5 is prohibited)
- 4: Core conclusions basically correct, minor omissions not affecting direction
- 3: Conclusions may be correct but evidence insufficient, or significant omissions
- 2: Conclusions may be wrong, re-analysis needed
- 1: Conclusions clearly wrong

[Mandatory GATE Tag]
Last line of report must be one of (on its own line):
GATE:bcheck:PASS
GATE:bcheck:FAIL:{score}

Rule: score >= 4.0 -> PASS, otherwise -> FAIL:{score}

[Mandatory Summary]
After review, generate a one-screen summary:
- Whether core conclusions hold (per-item)
- Reviewer's own verification conclusions
- Missed issues (if any)
- Score + specific issues
- Mark summary with separator lines

[Report Saving]
Write the complete report to the specified output path.

Key Points

• Actual execution verification is mandatory — reviewer must curl APIs, grep code, query DBs to get their own evidence
• Reviewer attitude is aggressive — "your analysis may be completely wrong"
• The document itself must contain actual evidence; bcheck is the second pair of eyes, not a replacement for the executor's analysis
• Score >= 4 to pass (passing doesn't mean no revisions needed)
• 1 initial review + up to 2 re-reviews max
• Treat every review as the last one — give it your all

Review Flow

Executor Agent flow:

1. Complete analysis or proposal, run verification commands for evidence
2. Write analysis document (path of your choice)
3. Call an independent review Agent for bcheck, passing complete document + report output path
4. On receiving the report:
   • PASS -> Append ## bcheck Response to your document, address each item: [Adopted] / [Not Adopted] + reasoning
   • FAIL -> Fix the analysis -> call bcheck again
5. Maximum 3 rounds (1 initial + up to 2 re-reviews)

---

This skill is part of the Parallax framework. Visit the project page for the full Parallax Loop explanation and latest skills. More AI4E tips: ai4e.dev

---

Source: AI4MSE/Parallax — distributed by TomeVault.