Plan Critic

Plan Critic (Copilot)

Critique an implementation plan before any code is written. The cost of revising a plan is near-zero; the cost of revising executed code is hours of rework. This skill keeps orchestration in-session: the current Copilot agent triages the plan, builds one Grounding Brief from the codebase, delegates to three native persona reviewer agents, and synthesizes a single verdict. Nothing is approved until three independent personas have inspected the plan against ground truth.

Core principle: A plan that names files generically has not been read. A plan that references verify_jwt_token at auth/middleware.go:42 has been read. The job of this skill is to tell those apart and force the second.

The reviewers (3 native agents)

Bundled as custom agents under agents/ (namespaced, e.g. plan-critic:skeptic-reviewer). Their exact instruction blocks live in [references/personas.md](references/personas.md) - read it before Phase 3 so you brief each reviewer faithfully.

| Reviewer agent | Lens | |---|---| | skeptic-reviewer | What's missing - hidden assumptions, edge cases, failure modes, rollback gaps, missing verification criteria | | architect-reviewer | Is it built right - file selection, execution order, convention fit, scope, backward compatibility, system impact | | verifier-reviewer | Did the author actually read the code, or skim file names - symbol specificity, cross-check vs the Grounding Brief, architectural awareness |

Keep this reviewer framing internal (see Privacy / scope).

Parsing the request

The user's text (or an attached file / pasted block) is the plan to review. Resolve it first: read the file if a path was given, use the inline text if pasted, or adopt the most recent plan produced in this conversation. If no plan is resolvable, ask which plan to review - do not invent one.

Workflow

Phase 1 - Triage (fast, do not delegate)

Resolve the plan input, then run a fast scope check. This phase never delegates - if it finds a fundamental issue, report it and stop.

1. Is this actually a plan? It must describe what will change and where, not just narrative discussion. If it has no concrete file/function/step list, ask the user to convert it into a plan first, then stop.
2. Plan size scan. Count files mentioned, count distinct steps/phases, and note whether line numbers or symbol names appear.
3. Trivial-change escape hatch. If the plan describes a change that fits in one sentence (typo, single-line fix, rename in one file), tell the user planning overhead is not warranted - recommend executing directly - and stop. Do not delegate to reviewers for trivial plans.
4. Scope warning. If the plan touches 7 or more files, surface this immediately: context degrades plan quality past this threshold and reviewers begin missing cross-file interactions. Recommend splitting into sub-plans before review, or proceeding only with explicit acknowledgment.

If triage finds a fundamental issue (not a plan, trivial, or oversized), report and stop here.

Phase 2 - Grounding Brief (build INLINE, do not delegate)

Build the Grounding Brief yourself in the orchestrator by grep/read over the codebase. All three reviewers share this one brief - do not delegate this phase and do not let each persona re-explore.

Produce, against the real code:

1. File verification - for every file the plan names, confirm it exists; list missing/misnamed paths.
2. Symbol verification - for every function/type/symbol the plan names, locate it (path:line) and flag any that do not exist or are misnamed (e.g. plan says UserStore.Save, code has UserRepository.Persist at store/user.go:118).
3. Caller blast radius - for functions the plan modifies, count callers and list 3-5 locations.
4. Existing patterns - find 2-3 existing examples of the kind of change proposed (e.g. existing handlers a new handler should mirror).
5. Related tests - find tests covering the modules being changed, and note coverage gaps.

Grounding Brief shape:

## Grounding Brief
### File verification
### Symbol verification
### Callers (blast radius)
### Existing patterns
### Related tests

The Brief is load-bearing - pass it verbatim to every reviewer. Trust the Brief over the plan's claims: when the plan says X exists and the Brief says X not found, the Brief wins.

Phase 3 - Reviewer delegation (sequential by default)

Read [references/personas.md](references/personas.md), then delegate each reviewer through the agent tool, giving each one: the full plan text, the Grounding Brief from Phase 2, that persona's mandate, and the instruction that its first response line must be ## review. Delegate the Skeptic, then the Architect, then the Verifier.

Default to sequential reviewer delegation - one agent at a time. Copilot parallel fan-out over-parallelizes and burns AI credits and offers no reliability guarantee; sequential is the reliable path. Only fan out in parallel if the user explicitly asks, then supervise with listagents/readagent and confirm each returned a payload before synthesizing.

Validate each reviewer result: it must start with ## review and contain the persona's contract sections. If a result is malformed or empty, re-delegate that reviewer once with the same plan + Brief; if still malformed, continue and note the missing lens in the synthesis. Never surface raw reviewer envelopes or readiness text.

Phase 4 - Synthesis and verdict

Recall the Phase 1 scope flags and the Phase 2 Grounding Brief before reading reviewer output, so the verdict stays anchored to verified evidence rather than persona assertion. Then merge the three reviews into one verdict - do not concatenate them:

1. Deduplicate - if two reviewers flagged the same missing file or symbol, merge into one finding.
2. Separate blockers from improvements - blocking issues first, then refinements, then nits.
3. Surface disagreement, don't flatten it - if one reviewer leans APPROVE and another REJECT, present the conflict explicitly and let the user adjudicate; trust the more conservative reviewer when forced to pick.
4. Cross-cutting insight - call out findings that emerge only from combining perspectives (e.g. Verifier says a file was not read + Architect says the plan ignores its 20 callers + Skeptic says no test covers the new behavior = one strong blocker).
5. Pick one verdict:

• APPROVE - strategy sound, depth adequate, no critical gaps. Safe to execute.
• REFINE - on the right track but needs specific, concrete additions/corrections first. Provide an ordered refinement list (if it grows past ~5 items, switch to REJECT).
• REJECT - fundamental issues (wrong approach, missed key files, hallucinated symbols, surface reading). Must be re-planned after the author reads more code.

Return findings only - do not rewrite the plan yourself; evaluation and authorship are different jobs.

Synthesis output shape

# Plan Review

## Verdict: [APPROVE | REFINE | REJECT]
**One-line summary:** 
**Plan stats:** 

## Triage notes

## Grounding Brief

## Findings
 refinements -> nits>

## Disagreement

## Refinement list

## Rejection rationale

## Next action
 and re-plan from scratch.">

Privacy / scope

The reviewer personas are internal review aids. Keep reviewer framing internal - present one synthesized verdict to the user, not three raw envelopes; if a finding leaves the team, restate it in your own voice. This skill reviews a plan, before code exists - it does not execute the plan, and it is not a substitute for reviewing an already-written diff (use a staff-level or language-specific code review for that).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: smykla-skalski
• Source: smykla-skalski/sai
• License: MIT

Install and usage instructions live in the source repository linked above.