draft-plan

name	draft-plan
disable-model-invocation	false
argument-hint	[output FILE] [rounds N] <description...>
description	Draft a high-quality plan through iterative adversarial review. Multiple rounds of research, drafting, review, devil's advocate, and refinement until the plan converges. Output is a plan file ready for /run-plan. Usage: /draft-plan [output FILE] [rounds N] <description...>

Codex Compatibility

This block applies only to the Codex-installed copy of this skill. It is an adapter layer for the original Claude slash-command instructions, not a replacement for the upstream workflow.

Invocation: in Codex, invoke this skill by naming it, for example run-plan ..., or by using the original slash command when the user supplied it. Treat $ARGUMENTS as the text after the skill name or slash command. Pass only the intended argument tail when one ZSkill calls another.

Tool mapping: if Codex exposes a subagent tool, map Claude Agent, Task, or subagent dispatch to that tool. Exploration-only work uses an explorer-style agent; implementation work uses a worker-style agent with explicit file/module ownership and instructions not to revert others work; review and devil-advocate work uses a general review agent. If no subagent tool is available, run the subtask inline, clearly label the degraded independence, and do not claim fresh-agent isolation. Read, Grep, Glob, and Bash map to shell reads, rg, and command execution; manual Edit/Write operations map to patch-based edits.

Skill calls: when this skill invokes another ZSkill such as /run-plan, /draft-plan, /verify-changes, or /commit, load and follow the selected skill instructions from Codex's available skills. If you must read the file directly, prefer $PROJECT_ROOT/.agents/skills/<skill>/SKILL.md, then $HOME/.agents/skills/<skill>/SKILL.md, then legacy $HOME/.codex/skills/<skill>/SKILL.md. Do not recursively re-enter the same skill unless the workflow explicitly requires it. Preserve ZSKILLS_PIPELINE_ID and related tracking environment across skill boundaries.

Tracking: tracking files belong under the main repository root at .zskills/tracking/, never under ~/.codex/skills. Resolve the main root with the original git rev-parse --git-common-dir pattern. Keep .zskills/tracking/ ignored. Do not delete or clear tracking except through an explicit user-requested clear-tracking workflow.

Landing modes: preserve direct, cherry-pick, and pr. Explicit direct, pr, or cherry-pick arguments win and should be stripped before downstream phase parsing. Otherwise read .codex/zskills-config.json, then .claude/zskills-config.json, then fall back to cherry-pick. If both config files exist and disagree on landing or main protection, stop before landing and report the conflict. locked-main-pr remains the preset name for PR mode with main protection.

Foreground runner bridge: Claude cron tools (CronList, CronCreate, CronDelete) are not the Codex implementation for run-plan finish auto. In Codex, finish auto must use a visible foreground parent runner that stays attached to the initiating REPL and launches one fresh codex exec child chunk per phase. Prefer scripts/zskills-runner.sh, then $PROJECT_ROOT/.agents/skills/scripts/zskills-runner.sh, then $HOME/.agents/skills/scripts/zskills-runner.sh, then legacy $HOME/.codex/skills/scripts/zskills-runner.sh. If the runner is unavailable, execute at most one phase, write the normal report/tracking handoff, and do not claim autonomous completion. Child prompts containing RUNNER-MANAGED CHUNK must not invoke the runner again; they execute exactly one incomplete phase and stop after durable plan/report/tracking evidence.

Hook fallback: Claude hooks are not enforced by Codex in this environment. Compensate with inline preflight checks before commits, cherry-picks, PR merge/auto-merge, worktree deletion, or tracking cleanup: inspect status, protect unrelated changes, verify branch/mode, and preserve .zskills/tracking.

Helper scripts: generated Codex installs include shared helpers under .agents/skills/scripts/ for project installs or $HOME/.agents/skills/scripts/ for user installs. Prefer project-local helpers at $PROJECT_ROOT/scripts/ when present, otherwise use the installed helper path, with $HOME/.codex/skills/scripts/ as a legacy fallback. zskills-runner.sh, zskills-gate.sh, and zskills-post-run-invariants.sh are Codex foreground-runner helpers. If a required helper is unavailable, use the explicit fallback instructions in the skill and report the degraded procedural path.

See ~/.codex/skills/ZSKILLS_CODEX_INTEGRATION.md for the shared adapter contract.

/draft-plan [output FILE] [rounds N] <description...> — Adversarial Plan Drafter

Produces high-quality plan documents through iterative adversarial refinement. Multiple agents research, draft, review, poke holes, and refine until the plan is solid enough to execute with /run-plan.

The insight: /run-plan executes a plan faithfully — so plan quality IS output quality. Investing in adversarial plan refinement pays off massively downstream. A weak plan executed perfectly is still a weak result.

Ultrathink throughout.

Arguments

/draft-plan [output FILE] [rounds N] <description...>

• output FILE (optional) — where to write the plan. Default: plans/<slug-from-description>.md
• rounds N (optional) — max review/refine cycles. Default: 3. The process exits early if a round converges (no substantive new issues).
• description (required) — everything after the recognized keywords. Can be brief ("add dark mode") or detailed ("implement thermal domain with conduction, convection, radiation, and multi-domain coupling to electrical via Joule heating").

Detection: scan $ARGUMENTS from the start for recognized patterns:

• output followed by a path — explicit output file
• The first token ending in .md — output file (only when it's the first argument, before the description starts). This avoids false positives on description words like README.md or CLAUDE.md. If the token contains /, use as-is; otherwise prepend plans/.
• rounds followed by a number — max review cycles
• Everything else (from the first unrecognized non-flag token onward) is the description

Examples:

• /draft-plan Add dark mode to the editor
• /draft-plan THERMAL_PLAN.md Implement thermal domain → writes plans/THERMAL_PLAN.md
• /draft-plan plans/THERMAL_PLAN.md Implement thermal domain → same
• /draft-plan output plans/THERMAL_PLAN.md rounds 5 Implement thermal domain
• /draft-plan rounds 5 Implement thermal domain with multi-domain coupling
• /draft-plan Fix the README.md formatting → description only, no output file detected

Pre-check — Existing file

If the output file already exists, read it first. The old plan IS research input — it contains the original intent, structure, and possibly partial progress. Tell the research agents: "An existing plan file exists at <path>. Read it and incorporate its intent. This is a modernization, not a blank-slate rewrite." The adversarial review should check that no intent from the original was lost.

This handles the common case of modernizing old-format plans: /draft-plan plans/OLD_PLAN.md Modernize with progress tracker and phases

Phase 1 — Research (parallel agents)

Tracking fulfillment

Determine the tracking ID: use the ID passed by the parent skill if this is a delegated invocation, or derive from the output file slug if standalone (e.g., plans/FEATURE_PLAN.md → feature-plan). Create the fulfillment file in the MAIN repo. The pipeline subdir follows Option B layout: when delegated, parent exports ZSKILLS_PIPELINE_ID via env and we write into the parent's subdir; when standalone, we own our own subdir:

MAIN_ROOT=$(cd "$(git rev-parse --git-common-dir)/.." && pwd)
PIPELINE_ID="${ZSKILLS_PIPELINE_ID:-draft-plan.$TRACKING_ID}"
mkdir -p "$MAIN_ROOT/.zskills/tracking/$PIPELINE_ID"
printf 'skill: draft-plan\nid: %s\noutput: %s\nstatus: started\ndate: %s\n' \
  "$TRACKING_ID" "$OUTPUT_FILE" "$(TZ=America/New_York date -Iseconds)" \
  > "$MAIN_ROOT/.zskills/tracking/$PIPELINE_ID/fulfilled.draft-plan.$TRACKING_ID"

Research agents

Codex override: dispatch the agents below with Codex subagents only when a subagent tool is available. If no subagent tool is available, run each research focus inline as separate labeled passes, persist the same research summary files, and report that the plan used inline research rather than parallel fresh-agent research.

Dispatch multiple Explore agents in parallel to investigate the problem space. Each agent gets the full description and a specific research focus:

1. Codebase agent — find all relevant source files, understand current architecture, identify what exists and what needs to change. Map dependencies, shared infrastructure, and potential conflicts.

2. Patterns agent — read existing plans in plans/ for format and style reference. Read CLAUDE.md for constraints (no external solvers, no bundlers, etc.). Identify conventions and rules that the plan must respect.

3. Domain agent — research the specific technical domain. For solver work: understand the math, numerical methods, stability requirements. For UI work: understand the interaction model, accessibility, existing patterns. For infrastructure: understand the deployment, CI, hosting.

4. Prior art agent — check git history for related work, previous attempts, known issues. Read any existing plan files or progress docs that overlap with the description. Find relevant GitHub issues.

Consolidate the research into a single summary and write it to a file (e.g., /tmp/draft-plan-research-<slug>.md). The <slug> comes from the output filename if one was provided (e.g., FEATURE_EXPORT → /tmp/draft-plan-research-FEATURE_EXPORT.md). If no output file was given, derive from the description. Do not rely on keeping the research in memory — context compaction will degrade it across multiple rounds of adversarial review. The file persists through all phases.

The summary should cover:

• What exists today (relevant code, architecture)
• What needs to change or be built
• Constraints and rules that apply
• Prior art and lessons learned
• Open questions or uncertainties

Similarly, write each round's review findings to files:

• /tmp/draft-plan-review-round-N.md — reviewer + devil's advocate findings
• Pass these file paths to the refiner agent so it has the full context

Scope check — is this too big for one plan? After consolidating research, assess whether you can write a plan where every phase has specific, implementable work items and testable acceptance criteria, in roughly 6 or fewer phases. The question is not "can I list phases" but "can I spec each phase precisely enough that an implementing agent won't have to guess?"

Signs the task is too big for one plan:

• You'd need 8+ phases to cover the scope properly
• Phases would be vague ("implement thermal domain") because the breadth prevents precise specs for each one
• The task contains 2+ sub-problems that share no files or infrastructure
• You find yourself thinking "Phase 4 will probably need to be broken into sub-phases at implementation time"

Past failure: a Dashboard Blocks plan covered too much — phases were vague, the implementing agent took shortcuts, and each phase had to be broken into 5-10 sub-phases at execution time with no adversarial review.

Common pattern: tasks that add multiple block types alongside new engine infrastructure (new solver, new domain, new codegen path) should decompose into infrastructure-first plan + block additions via /add-block delegate phases after the engine lands.

If the task is too big, present this to the user:

This task is too broad for one well-specified plan. It decomposes into N sub-problems: [list them with one-line descriptions]. I recommend using /research-and-plan to handle the decomposition and draft focused sub-plans for each. Proceed?

On approval, invoke /research-and-plan with:

• The original description (full, unmodified)
• The research file path (/tmp/draft-plan-research-<slug>.md) Then exit — do not proceed to Phase 2. /research-and-plan takes over from here.

If the user says no (e.g., "just plan the first part"), narrow the scope per their feedback and proceed to Phase 2 with the focused scope.

Post-research tracking

After consolidating research into the summary file, create the research step marker:

MAIN_ROOT=$(cd "$(git rev-parse --git-common-dir)/.." && pwd)
PIPELINE_ID="${ZSKILLS_PIPELINE_ID:-draft-plan.$TRACKING_ID}"
printf 'completed: %s\n' "$(TZ=America/New_York date -Iseconds)" \
  > "$MAIN_ROOT/.zskills/tracking/$PIPELINE_ID/step.draft-plan.$TRACKING_ID.research"

Present the research summary to the user. If running interactively (user invoked /draft-plan directly), wait for input:

Research complete. Summary written to /tmp/draft-plan-research-<slug>.md. Here's the overview: [brief summary] [Scope check result — either "this fits in one plan" or the decomposition recommendation above]

Anything to add, correct, or steer before I draft the plan?

Incorporate any user feedback, then immediately proceed to Phase 2. Do not stop here. The checkpoint is a pause for steering, not the end of the skill. After the user responds (even if they just say "looks good" or "continue"), move to Phase 2 without being asked again.

If running as a subagent (dispatched by /research-and-plan or /research-and-go), skip the user checkpoint — proceed directly to Phase 2. The decomposition was already approved by the user in the parent skill.

Phase 2 — Draft

A single agent produces the initial plan based on the consolidated research and user feedback. The plan MUST follow a format that /run-plan can consume:

Landing mode hint

Before writing the plan, determine which landing mode hint (if any) to embed near the top. Resolution order:

1. Explicit description suffix — if the description passed to /draft-plan ends with Landing mode: pr or Landing mode: direct (as appended by /research-and-plan), that wins. Strip this suffix from the description before using it in the plan body.

2. Config default — otherwise read .claude/zskills-config.json for execution.landing:

   PROJECT_ROOT=$(cd "$(git rev-parse --git-common-dir)/.." && pwd)
   CONFIG_FILE="$PROJECT_ROOT/.claude/zskills-config.json"
   LANDING_HINT=""
   if [ -f "$CONFIG_FILE" ]; then
     CONFIG_CONTENT=$(cat "$CONFIG_FILE")
     if [[ "$CONFIG_CONTENT" =~ \"landing\"[[:space:]]*:[[:space:]]*\"([^\"]*)\" ]]; then
       LANDING_HINT="${BASH_REMATCH[1]}"
     fi
   fi
   

3. Fallback — if neither the description suffix nor the config specifies a mode, treat it as cherry-pick (the default) — no hint emitted.

Based on the resolved mode, prepend one of the following blockquotes immediately after the # Plan: <Title> heading and before ## Overview:

• pr:

  > **Landing mode: PR** -- This plan targets PR-based landing. All phases
  > use worktree isolation with a named feature branch.
  

• direct:

  > **Landing mode: direct** -- This plan targets direct-to-main landing.
  > No worktree isolation.
  

• cherry-pick or absent: no blockquote (default behavior — do not emit a hint).

This is a hint, not enforcement. The hint exists so the implementing agent (and any human reader) knows which landing model the plan was drafted for. At execution time the /run-plan argument (pr, direct, or unspecified) always takes precedence — the embedded hint never overrides an explicit /run-plan flag.

Required plan structure

Every plan file MUST begin with YAML frontmatter so that /run-plan can track metadata (especially which GitHub issue to close on completion):

---
issue: N          # GitHub issue number (omit if not created from an issue)
title: Plan Title
created: YYYY-MM-DD
status: active    # active | complete
---

Frontmatter rules:

• issue — include ONLY when /draft-plan is invoked from /fix-issues plan (or any context that supplies a GitHub issue number). When invoked standalone with no issue context, omit the issue: field entirely.
• title — the plan title. Because the frontmatter is the authoritative reference, do NOT duplicate the issue number in the # Plan: <Title> heading.
• created — use the current date (YYYY-MM-DD).
• status — always starts as active. /run-plan updates this to complete when all phases finish, which also signals it to close the linked GitHub issue (if one is present).

# Plan: <Title>

> **Landing mode: PR** -- This plan targets PR-based landing. All phases
> use worktree isolation with a named feature branch.
<!-- OR, if direct mode:
> **Landing mode: direct** -- This plan targets direct-to-main landing.
> No worktree isolation.
-->
<!-- Omit the blockquote entirely when landing mode is cherry-pick or unset. -->

## Overview
Brief description of what this plan accomplishes and why.

## Progress Tracker
| Phase | Status | Commit | Notes |
|-------|--------|--------|-------|
| 1 — <name> | ⬚ | | |
| 2 — <name> | ⬚ | | |
| ...         | ⬚ | | |

## Phase 1 — <Name>

### Goal
One-sentence summary of what this phase accomplishes.

### Work Items
- [ ] Item 1 — specific, implementable description
- [ ] Item 2 — ...

### Design & Constraints
Verbatim specs: formulas, state equations, algorithms, data structures,
API contracts. Everything an implementing agent needs to write the code
without guessing. If there are formulas, write the formulas. If there are
constraints, state them explicitly.

### Acceptance Criteria
- [ ] Criterion 1 — specific, testable condition
- [ ] Criterion 2 — e.g., "test free vibration x(t) = A cos(ωt) within 1e-6"

### Dependencies
Which phases must be complete before this one can start.

## Phase 2 — <Name>
...

Drafting rules

• Be specific, not vague. "Implement the Spring component using F = -kx with internal displacement state" — not "implement springs."
• Include formulas and algorithms verbatim. The implementing agent will have ONLY the plan text. If the formula isn't in the plan, the agent will guess it wrong.
• Acceptance criteria must be testable. "Works correctly" is not testable. "Test produces output matching analytical solution within 1e-6 tolerance" is testable.
• Phase boundaries at natural breaks. Each phase should be independently implementable and verifiable. Don't put shared infrastructure in Phase 3 if Phases 1 and 2 need it.
• Size phases for agent execution. Each phase should be completable by one agent in under 2 hours: implement, test, verify, commit. Scope by concrete units (components, files, features) — NOT by time estimates. A phase with ~3-5 components and ~500 lines of new code is right. A phase with 15 components or 2000+ lines should be split. Do NOT use time estimates ("this will take 2 weeks") — LLMs cannot estimate time. Use scope: number of components, files touched, test count.
• Estimate scope honestly. If a phase has 7 components, list all 7. Don't say "implement key components" and hope the agent picks the right ones.
• Dependencies must be explicit. If Phase 3 needs the base class from Phase 1, say so.

Phase 3 — Adversarial Review (parallel)

Dispatch two agents simultaneously against the current draft. Each gets the full draft AND the research summary from Phase 1.

Reviewer agent

Checks for completeness, correctness, and feasibility:

• Are all work items specific enough to implement without guessing?
• Do acceptance criteria cover every work item?
• Are dependencies correct and complete?
• Is the phase ordering optimal? (shared infrastructure first?)
• Are there missing phases? (tests? documentation? integration?)
• Is the scope realistic per phase? (1000+ lines each is the norm)
• Does the plan respect all constraints from CLAUDE.md and research?
• Are formulas and algorithms correct?
• Will /run-plan be able to parse phases and status from this format?

Evidence discipline. When a finding makes an empirical claim (a file, function, tool, or library has/lacks property X), include a concrete Verification: line — the exact file:line, grep, schema quote, or command output that reproduces the evidence. The refiner will re-run these checks before acting. Structural/judgment findings don't need a reproducer but should say Verification: judgment — no verifiable anchor explicitly. Never write an empirical-sounding claim without something the refiner can independently re-check.

Devil's advocate agent

Genuinely adversarial — tries to find ways the plan will fail:

• Wrong assumptions — what does the plan assume that might not be true? ("Assumes the solver handles stiff systems" — does it actually?)
• Missing edge cases — what will break at implementation time that the plan doesn't address? ("What happens when the spring constant is zero?")
• Deferred hard parts — are the difficult items buried in later phases or hidden behind vague language? ("Phase 4: integrate everything" — that's where all the complexity is.)
• Hidden dependencies — does Phase 3 actually need something from Phase 2 that isn't listed as a dependency?
• Overly optimistic scope — is "implement 7 components" in one phase actually feasible? Should it be split?
• Specification gaps — "when an agent tries to implement Phase 2b using ONLY the text in this plan, what will go wrong? What will it have to guess?"
• Integration risks — will phases that work in isolation break when combined? What integration testing is missing?
• Constraint violations — does any phase implicitly require an external library, a build step, or something else CLAUDE.md prohibits?

The devil's advocate must produce specific, actionable findings — not generic concerns. "Phase 3 might be complex" is useless. "Phase 3 says 'implement the DAE solver' but doesn't specify the index reduction method, matrix factorization approach, or convergence criteria — the agent will have to guess all three" is actionable.

The same evidence discipline from the reviewer section applies: every empirical claim needs a Verification: line. The devil's advocate is especially prone to generating plausible-sounding-but-false claims because its job is pattern-generating failure modes, not verifying them. Discipline is load-bearing here — the refiner will re-check, and claims whose evidence doesn't reproduce will not drive fixes.

Post-review tracking

After both reviewer and devil's advocate agents return their findings, create the review step marker:

MAIN_ROOT=$(cd "$(git rev-parse --git-common-dir)/.." && pwd)
PIPELINE_ID="${ZSKILLS_PIPELINE_ID:-draft-plan.$TRACKING_ID}"
printf 'round: %s\ncompleted: %s\n' "$ROUND" "$(TZ=America/New_York date -Iseconds)" \
  > "$MAIN_ROOT/.zskills/tracking/$PIPELINE_ID/step.draft-plan.$TRACKING_ID.review"

Phase 4 — Refine

A single agent receives:

• The current draft
• The reviewer's findings
• The devil's advocate's findings

Verify-before-fix (mandatory)

Before touching the draft, the refiner must attempt to reproduce the cited evidence for each finding. The reviewer/DA produce hypotheses; the refiner is the gate that tests them against reality.

For each finding with an empirical claim:

1. Read the Verification: line and run its check (Read the file, run the grep, check the schema, run the command).
2. Record outcome: Verified (evidence reproduces → act on finding), Not reproduced (evidence does not match reality → justify, do NOT fix based on this finding), or No anchor (empirical-sounding claim without verifiable citation → scrutinize, either locate an anchor yourself or justify-not-fix).

Judgment findings (marked Verification: judgment) skip step 1 and go straight to fix-or-justify on merit.

Why this exists. Devil's-advocate findings are by role generated to be plausible failure modes, not verified truths. Past failure: a DA claimed a tool's input lacked a field, citing a tangentially-related file; the refiner accepted and rewrote a whole phase on a false premise. A 30-second check of the tool schema would have caught it. Verify-before-fix is the gate that turns "the DA said so" into "I checked."

Address every finding

It produces an improved draft that addresses every finding. For each finding, it must either:

1. Fix it — update the plan to resolve the issue
2. Justify — explain why it's not actually a problem (with evidence, including the "evidence did not reproduce" and "claim not verifiable" cases from the verify-before-fix block)

It may NOT ignore findings or defer them. The refiner's output is the new draft for the next round.

The refiner's output must include a disposition table listing each finding with an Evidence column (Verified / Not reproduced / No anchor / Judgment) and the disposition (Fixed / Justified + reason).

Phase 5 — Convergence Check

After each round of review + refinement:

1. Count substantive issues — how many findings from the reviewer and devil's advocate were real problems (not false positives)?

2. Check convergence:

   • 0 substantive issues → converged. Proceed to Phase 6.
   • Substantive issues remain AND rounds < max → back to Phase 3 with the refined draft.
   • Max rounds reached → proceed to Phase 6 with a "remaining concerns" section noting unresolved issues.

3. Track round history — keep a log of each round's findings and resolutions. This goes into the final plan's quality section.

Phase 6 — Finalize

1. Add a Plan Quality section to the end of the plan:

   ## Plan Quality
   
   **Drafting process:** /draft-plan with N rounds of adversarial review
   **Convergence:** [Converged at round M / Max rounds reached]
   **Remaining concerns:** [None / List of unresolved issues]
   
   ### Round History
   | Round | Reviewer Findings | Devil's Advocate Findings | Resolved |
   |-------|-------------------|---------------------------|----------|
   | 1     | 5 issues          | 7 issues                  | 12/12    |
   | 2     | 2 issues          | 3 issues                  | 5/5      |
   | 3     | 0 issues          | 0 issues                  | Converged|
   

2. Write the plan file to the output path. The user can't review what they can't read — plans are often too large to meaningfully summarize in chat. Write first, then let the user read the actual file.

3. Update the plan index:

   • If plans/PLAN_INDEX.md exists, add a row to the "Ready to Run" table with the new plan's filename (as a relative link), phase count, first phase name, priority Medium, and a one-line note from the overview.
   • If plans/PLAN_INDEX.md does not exist, include in the report:

     Run /plans rebuild to generate a plan index.

4. Present the result:

   Plan drafted in N rounds (converged / max rounds reached). [Remaining concerns if any]

   Written to plans/THERMAL_PLAN.md — open it up and let me know what to change.

   Execute with: /run-plan plans/THERMAL_PLAN.md Or with scheduling: /run-plan plans/THERMAL_PLAN.md auto every 4h now

Post-finalize tracking

After writing the plan file and updating the index, create the finalize step marker and update the fulfillment file to complete:

MAIN_ROOT=$(cd "$(git rev-parse --git-common-dir)/.." && pwd)
PIPELINE_ID="${ZSKILLS_PIPELINE_ID:-draft-plan.$TRACKING_ID}"
printf 'completed: %s\n' "$(TZ=America/New_York date -Iseconds)" \
  > "$MAIN_ROOT/.zskills/tracking/$PIPELINE_ID/step.draft-plan.$TRACKING_ID.finalize"

printf 'skill: draft-plan\nid: %s\noutput: %s\nstatus: complete\ndate: %s\n' \
  "$TRACKING_ID" "$OUTPUT_FILE" "$(TZ=America/New_York date -Iseconds)" \
  > "$MAIN_ROOT/.zskills/tracking/$PIPELINE_ID/fulfilled.draft-plan.$TRACKING_ID"

Key Rules

• The plan is the spec. Every implementing agent will have ONLY the plan text. If it's not in the plan, it doesn't exist. Write formulas, not descriptions of formulas. Write acceptance criteria, not hopes.
• Devil's advocate must be genuinely adversarial. Not a rubber stamp. The goal is to find real problems before /run-plan hits them at implementation time. A comfortable review is a useless review.
• Every finding must be addressed. The refiner cannot ignore or defer reviewer/devil's-advocate findings. Fix it or justify why it's not a problem.
• Verify findings before baking fixes. Reviewer and DA findings are hypotheses, not mandates. The refiner must reproduce each empirical claim before acting on it. A plausible-sounding claim with no verifiable anchor — or one whose evidence doesn't reproduce — is not a mandate to fix. Devil's advocate findings are especially prone to confidently-false empirical claims because the role incentivizes plausibility, not truth.
• Convergence means no new substantive issues. Not "the same issues rephrased." If the devil's advocate keeps finding real new problems, the plan isn't ready.
• Respect constraints. The plan must not require anything CLAUDE.md prohibits: no external solvers, no bundlers, no external dependencies without approval.
• Phase boundaries at natural breaks. Each phase must be independently implementable and verifiable by /run-plan.
• Progress tracker is mandatory. /run-plan needs it to track status. Start all phases as ⬚ (not started).
• Ultrathink throughout. Every agent in the process should use careful, thorough reasoning.

Edge Cases

• Description is very brief ("add dark mode") — the research phase does the heavy lifting, exploring what "dark mode" means for this codebase
• Description is very detailed (multi-paragraph brief) — research agents validate and expand rather than starting from scratch
• Plan converges in round 1 — great, write it. Don't force unnecessary rounds.
• Plan doesn't converge after max rounds — write it with the "remaining concerns" section. The user decides whether to proceed or refine further.
• User adds context after research — incorporate it into the draft. Don't ignore user input.
• Output file already exists — warn the user, ask before overwriting
• Plan is too large for one file — suggest splitting into multiple plan files, one per major phase group