agentleFS
DasDigitaleMomentumDasDigitaleMomentum/opencode-processing-skills/skills/review-implementation-plan/SKILL.md

review-implementation-plan

Skill60 starsChanged 34 days agoMITView on GitHub

Independent single-phase or batch review of implementation plans against scope, codebase reality, actionability, and cross-phase consistency. Produces per-phase structured reviews with severity-rated findings.

The file

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 7 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this skill when:

    line 22
  2. Distinctive:

    Do not use this skill to:

    line 29
  3. Distinctive:

    Needs Revision: Do not execute. Accept or explicitly reject each blocking finding.

    line 111
  4. Distinctive:

    Follow the canonical template headings and frontmatter keys.

    line 123
  5. Distinctive:

    Use current code evidence for material feasibility or minimality findings.

    line 125
  6. Very specific:

    Use Major Gaps when missing gated intent or an unresolved blocking decision prevents a defensible technical plan.

    line 136
  7. Distinctive:

    Do not fan out one reviewer per phase by default or reconstruct each authoring pass.

    line 146
  8. Distinctive:

    Do not modify the implementation plan during review — only produce the review artifact.

    line 148
  9. Very specific:

    Do not discard the reviewer task_id until the primary has decided whether remediation is needed.

    line 149
  10. Very specific:

    Ensure the reviews/ directory exists before delegating (create if needed).

    line 150

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Skill: Review Implementation Plan

This skill provides an independent quality gate for implementation plans authored via author-and-verify-implementation-plan.

A reviewer fresh from the authoring context evaluates one implementation plan or an ordered batch against phase scope, existing code, and actionability criteria. Fresh independence does not mean starting a cold reviewer for every phase. Each review remains a per-phase plan artifact.


When to Use

Use this skill when:

  • An implementation plan has been authored/verified and you want to validate it before execution.
  • Multiple implementation plans have been authored sequentially and should be reviewed together as a dependency-ordered batch.
  • You want to confirm that the implementation plan is concrete enough for execute-work-package.
  • The user explicitly requests an implementation plan review.

Do not use this skill to:

  • Review the plan itself (use review-plan).
  • Review a completed implementation (use review-implementation).
  • Rewrite the implementation plan during the independent review pass. After review completion, the Primary may accept defects and explicitly instruct technical implementation-plan remediation through review-fix in the eligible source session. Conversation-owned plan corrections remain with update-plan.

Review Focus

The default priority is the smallest complete implementation: confirm that phase obligations are covered and no implementation step or new artifact is unauthorized, unnecessary, or needlessly indirect. Use the actual codebase to substantiate material concerns, but record only evidence-backed exceptions. The primary may add a focus via {{focus}}.

Review posture

Detect existing gold-plating without becoming an adversarial reviewer. A reduction finding must name a concrete planned step/artifact and show missing authorization, missing present necessity, or a smaller path in the current codebase that remains complete: the requested behavior must still work, affected real paths must still integrate, and applicable existing invariants must be preserved. Scope discipline is not permission to omit necessary work. Do not invent ideal architecture, hardening, tests, infrastructure, policy, or replacement work. Zero findings remains valid.

Testing, references, Reality Check wording, documentation, and other formal criteria are not checklist obligations. Report them only when a concrete defect would block or misdirect execution.


Execution Model

Roles

  • Primary (maintainer)
    • Invokes the review skill.
    • Delegates to delegate-strong (default) or general (for same-model perspective).
    • Receives review summary and decides on follow-up actions.
    • Routes a batch through one fresh reviewer session by default and retains that task_id for possible remediation.
  • Subagent (delegate-strong / general)
    • Starts without authoring context, then retains shared review context across a batch.
    • Uses the actual codebase selectively when a material concern needs evidence.
    • Writes the existing review artifact for every reviewed phase at plans/<name>/reviews/impl-plan-review-phase-N.md.
    • Reports material cross-phase conflicts in the relevant artifact and returns one aggregate digest.

Why delegate-strong (not doc-explorer)

Same rationale as review-plan: the reviewer must be independent from the authoring work. In batch mode that fresh reviewer keeps useful context between phases rather than repeatedly approaching each phase cold. delegate-strong provides the judgment depth needed to evaluate implementation feasibility and cross-reference plan claims against real code.

Routing Matrix (Who does what)

  • Writes: one or more existing plans/<name>/reviews/impl-plan-review-phase-N.md artifacts; no consolidated artifact is required.
  • Does NOT write: implementation plans, phase docs, or any other plan artifact.
  • Primary: owns the decision of whether to act on findings.
  • delegate-strong/general: performs the review, including codebase verification.

Workflow

1) Select review mode and prepare references

Primary gathers:

  • Review focus from the delegation prompt
  • plans/<name>/plan.md
  • One phase/implementation-plan pair for single-phase mode, or all selected pairs in dependency order for batch mode
  • Relevant existing docs only when they reduce a material code lookup

2) Delegate

Primary delegates to delegate-strong (or general) using tpl-review-impl-plan-prompt.md.

Provide:

  • Review mode and the ordered plan, phase, and implementation-plan paths
  • Docs references (if available)
  • One review output path per phase: plans/<name>/reviews/impl-plan-review-phase-N.md
  • Review focus (freetext — what to prioritize)

In batch mode, one fresh reviewer processes phases sequentially, reuses shared evidence, writes each per-phase artifact, reports only material shared-interface conflicts, and returns one aggregate digest. Split the batch only when unrelated domains or practical context capacity require it.

3) Receive summary

For single-phase mode, the subagent returns the same fields for that phase. For batch mode, it returns one aggregate digest containing:

  • Overall and per-phase verdicts (Ready / Needs Revision / Major Gaps)
  • Overall and per-phase reduction flags
  • Aggregate finding count by severity
  • Top 3 actionable findings across the batch
  • Required next action

4) Act on findings

Primary decides:

  • Ready: Proceed to execute-work-package only when reduction is not required and no Critical/Major findings remain.
  • Needs Revision: Do not execute. Accept or explicitly reject each blocking finding. Explicitly instruct accepted technical implementation-plan reductions once through review-fix, reusing correct and sufficient source-session review/diagnosis context under that skill's eligibility and assignment rules.
  • New authoring pass: Use a fresh author-and-verify-implementation-plan session when retained context/output is faulty, insufficient, or unavailable, with a finding-source pointer as a hint, not inherited context. A changed objective/gated scope also requires new authoring; resolve new Primary decisions before dependent work. Conversation-owned plan changes go through update-plan.
  • Major Gaps: Discuss with user; potentially revise phase scope via update-plan.

The remediation digest closes the accepted pass. Do not automatically re-review or continue until zero findings. Escalate materially changed scope or new risk/uncertainty to the Primary; a fresh review requires an explicit Primary or user decision.


Output Contract

Each review artifact plans/<name>/reviews/impl-plan-review-phase-N.md MUST:

  • Follow the canonical template headings and frontmatter keys.
  • Include a clear assessment with verdict, reduction flag, and brief reasoning.
  • Use current code evidence for material feasibility or minimality findings; do not exhaustively revalidate every reference when no concern exists.
  • Report only exceptions; do not reproduce phase-coverage or per-step disposition tables.
  • Give every finding a stable ID, severity, evidence, and concrete action.
  • State No findings when the review finds no material problem.

In batch mode, record material cross-phase findings in the affected per-phase artifact. Do not create a mandatory consistency section or consolidated artifact.

Verdict rules:

  • Ready requires Reduction Required: No and zero Critical/Major findings.
  • Any executable step or new artifact without clear authorization or present necessity is at least Major and requires Needs Revision.
  • Use Major Gaps when missing gated intent or an unresolved blocking decision prevents a defensible technical plan.

Rules

  • Examine the actual codebase only as needed to judge feasibility and substantiate findings. Do not perform exhaustive path/symbol certification as a formal exercise.
  • The reviewer must begin fresh from the authoring context. In batch mode, retain review context across phases; fresh perspective does not require a cold reviewer per phase.
  • Support both single-phase and batch review without changing the per-phase artifact naming convention.
  • Batch review is sequential in dependency order and produces one aggregate digest. Cross-phase checks are limited to actual shared interfaces and dependencies.
  • Do not fan out one reviewer per phase by default or reconstruct each authoring pass.
  • Findings are advisory decisions, not automatic edits. The primary must accept or explicitly reject blocking findings before progression.
  • Do not modify the implementation plan during review — only produce the review artifact.
  • Do not discard the reviewer task_id until the primary has decided whether remediation is needed.
  • Ensure the reviews/ directory exists before delegating (create if needed).
  • Zero findings is valid. Report testing, rollback, security, deployment, documentation, or policy only when a concrete defect exists; do not invent requirements or infrastructure.
  • Pay particular attention to new layers, modules, interfaces, generic infrastructure, cleanup, and future-phase preparation, but report only concrete unnecessary work.
  • Once a review is invoked, Reduction Required: Yes or unresolved Critical/Major findings block progression until the primary remediates or explicitly rejects them with rationale.
  • One review plus one accepted remediation pass is the default bound. Never start an automatic review/fix/re-review loop.

Templates

  • tpl-impl-plan-review.md — Canonical compact per-phase review output
  • tpl-review-impl-plan-prompt.md — Primary → reviewer delegation prompt

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.