Skip to main content
GeneralCodename-Inc

spectre-plan_review

👻 | Independent multi-lens review of plan.md and/or tasks.md — finds overengineering, missing verification, hallucinated deps, weak references

Stars
142
Source
Codename-Inc/spectre
Updated
2026-05-28
Slug
Codename-Inc--spectre--spectre-plan-review
View on GitHubRaw SKILL.md

// install — copy + paste into any project

mkdir -p .claude/skills && curl -fsSL https://raw.githubusercontent.com/Codename-Inc/spectre/HEAD/plugins/spectre-codex/skills/spectre-plan_review/SKILL.md -o .claude/skills/spectre-plan-review.md

Drops the SKILL.md into .claude/skills/spectre-plan-review.md. Works with Claude Code, Cursor, and any agent that loads SKILL.md files from .claude/skills/.

plan_review

Purpose

Produce the smallest correct plan through one evidence wave, correctness review, then subtraction-first simplification.

Inputs

  • $ARGUMENTS: feature root/name or descendant plan.md; optional --auto-apply scope-safe, --orchestrated.
  • Require {FEATURE_ROOT}/specs/plan.md or route to spectre-create_plan. Authority: concepts/scope.md, specs/prd.md, specs/ux.md, then explicit task_context.md requirements. Reuse existing research.

Working Set

  • Resolve one managed FEATURE_ROOT for this work from explicit/current-thread evidence only (physical directory wins; never branch/recency/lifecycle/scans). If none is confirmed, including when the candidate path is occupied, standalone MUST first load and follow Skill(spectre-feature-root) through DONE; orchestrated calls escalate. Keep writes beneath it and pass it unchanged.
  • Reports: reviews/plan_correctness.md and reviews/plan_review.md. Resume a sole correctness report only when its post-edit hash matches plan.md; otherwise use absent canonical paths or timestamped siblings.

Outputs + DONE

  • plan_correctness.md: route/model metadata and hashes; evidence/unknowns; findings; retained constraints/tests; dispositions and resulting edits.
  • plan_review.md: correctness path/hash; deletions, collapses, reuse, deferrals, retained exceptions, test reductions, findings/dispositions, and structural Before → After.
  • Reports are deltas; never restate the plan.

DONE when both stages complete; research ran once at most; every mechanism traces to a current requirement, constraint, prerequisite, or verified fact; every new complexity boundary has a valid exception; the plan is smaller in mechanisms, surfaces, process, or tests—or proves no safe reduction; behavior and constraints survive; no correctness Blocker/High remains; scope changes are withheld; findings are disposed; hashes/write bounds pass; and direct-mode Verification is executable.

Method / guardrails

Canonical Scope Invariant: correct and simplify implementation means without narrowing, expanding, or reinterpreting Scope. Mark boundary changes Scope Change Required; never auto-apply them.

  1. Evidence. Identify unsupported material claims. If needed, use at most one each @spectre_finder, @spectre_analyst, and @spectre_patterns, in parallel, for citation-first evidence/unknowns only (≤1,000 tokens total).

  2. Correctness. Read references/correctness-review.md and send it verbatim to a fresh reviewer with: plan, Scope, task-context, and report paths/hashes; compiled evidence/unknowns; edit mode and write bounds; route metadata.

  3. Simplification. After correctness writeback, read references/simplification-review.md and send it verbatim to a second fresh reviewer with: corrected plan, Scope, correctness-report, and output-report paths/hashes; edit mode and write bounds; route metadata. Supply existing evidence; allow one cited anchor spot-check.

  4. Writeback. Reviewers write only their report and plan.md, with the report written before plan edits; scope/context/tasks/code remain immutable. --auto-apply scope-safe permits Blocker/High and unambiguous Medium edits; otherwise ask all|blockers|IDs|skip, then continue on the same route. Record addressed|skipped|unresolved|scope-change. Stop on unresolved correctness Blocker/High, scope change, unavailable writeback, or failed schema/hash/scope/Out-of-Bounds checks. The primary may normalize mechanics, never semantics.

  5. Route. Run each stage fresh at high effort (20-minute limit): Codex → Claude Code opus; Claude Code → Codex gpt-5.6-sol. Record stage/runtime/model/effort/route. If unusable, use one clean-context native @spectre_reviewer with the same template and context. A usable review is terminal.

Handoff

Return routes, correctness status, deletions/retained exceptions, structural delta, dispositions, reports, withheld scope changes, and updated plan in ≤1,000 tokens. Orchestrated calls return; standalone success recommends spectre-create_tasks or direct spectre-execute.

Escalate-If

Plan missing; a material claim remains unknowable; scope change is required; correctness Blocker/High remains; writeback is unavailable; or schema/hash/scope/Out-of-Bounds checks fail.