Code Refactor

Code Refactor Skill

Purpose: Produce executable refactoring reports for an existing codebase, using a multi-round, multi-agent methodology. Two modes: AUDIT (high-level — find & rank opportunities) and PLAN (deep-dive — produce verifiable per-target refactor specs).

When to Use This Skill

Trigger on any of:

• "Audit this codebase for refactor opportunities"
• "What should we refactor in ?"
• "Tech debt review", "code quality audit", "maintainability review"
• "Do a deep-dive refactor plan for <feature/module>"
• "I want to give a developer a refactor task they can execute end-to-end"
• User wants either a high-level prioritized list or a specific, executable refactor proposal

Do NOT use for:

• Bug fixes (the code is wrong; this skill is for code that works but should be better)
• Behavior changes / new features (refactor = same behavior, better internals)
• Renames or trivial reorganization (a chat reply suffices)
• Production incidents — fix first, refactor later

The Two Modes

This skill has two operating modes. Pick one at the start; never run both in the same invocation.

The two modes share the same multi-round methodology. They differ in what gets drafted and what the reviewers are looking for.

Typical sequence: user runs AUDIT → reviews ranked list → picks top 1–3 → user runs PLAN on those targets → executor developer follows each PLAN.

Core Principles

1. Behavioral parity is non-negotiable. A refactor changes internals, never observable behavior. Every PLAN must prove parity with a pre-baseline and a post-comparison. If behavior must change, that's a feature/fix, not a refactor — say so and stop.
2. Evidence binds every claim. "This is slow", "this is hard to maintain", "this is duplicated" — each must cite file paths and line numbers. No vibes-based recommendations.
3. The executor closes the loop. A PLAN's success is measured by whether the receiving developer can: capture baseline → do the work → re-run the same checks → produce a verdict. If the loop doesn't close, the PLAN failed.
4. Specific over abstract. "Extract to a helper" is not a plan. "Move lines 47–93 of src/foo.ts into a new src/lib/normalizeFoo.ts exporting normalizeFoo(input: Foo): NormalizedFoo — see proposed diff below" is a plan.
5. Right-size the report. A 5-file project does not need a platform audit. A subsystem PLAN does not need a full-stack ceremony. The methodology is fixed; depth adapts.
6. Multi-round, multi-perspective. Draft 1 → Dual Review → Draft 2 → Final Review → Draft 3. Two reviewer perspectives in parallel catch different categories of issues. Skipping rounds costs more downstream than it saves now.
7. Project conventions win. Read the project's own coding standards / dev guidelines first. The refactor must end up more aligned with the project's conventions, never less.
8. Defaults are decisions. Every customization in a PLAN (test command, baseline location, artifact path) gets a default the executor can run with. The executor never blocks.
9. The PLAN is also a runbook. Final section instructs the executor end-to-end, including how to produce the HTML artifact. The PLAN is what and how, not just what.

The Seven Phases (shared by both modes)

Phase 1: Scoping & Context Load           (mode, scope, project conventions, test infra)
Phase 2: Discovery — Draft 1              (parallel agents, one per facet or per target)
Phase 3: Dual Review                      (skeptic + Reviewer B, parallel)
Phase 4: Draft 2                          (apply ALL fixes, parallel where possible)
Phase 5: Final Review                     (single reviewer, hunt remaining issues)
Phase 6: Draft 3                          (final fixes, mark Final)
Phase 7: Delivery                         (HTML artifact + working memory + handoff)

Phases are identical in shape across AUDIT and PLAN. The contents of each phase differ — see the per-mode sections.

Phase 1 — Scoping & Context Load

Identical for both modes.

1.1 Lock the mode

• AUDIT or PLAN?
• If PLAN: which 1–3 targets? (max three; if user offers more, push back and ask them to pick)
• Scope: full project, or a subsystem path / set of paths?

If the user is ambiguous, ask in a single batched message. Never one question at a time.

1.2 Load project context

Always read, in this order:

1. Project's CLAUDE.md / agents.md (if present)
2. Project's development guidelines / coding standards document (in DB project documents, or in repo)
3. Any PRD or spec the project follows
4. The repo's README, package.json (or equivalent), test config

Identify and capture:

• Tech stack — language, framework, runtime, package manager
• Test infra — unit / integration / e2e commands, test framework, location of tests
• Build / typecheck / lint commands — exact strings
• Linter & formatter — which configs are authoritative
• Conventions — naming, file org, error model, validation library, logging shape
• Branch & commit conventions — Conventional Commits? squash policy? PR template?
• Deployment — what "shipped" looks like (CI? PM2? Docker?)
• Forbidden zones — anything the project explicitly says do-not-touch

Save this context to a working note before drafting — it gets referenced in every subsequent phase.

1.3 Decide the discovery split

For AUDIT, run the 7 default facets (drop any that don't apply); add extension facets if the project clearly needs them — see Phase 2 — AUDIT.

For PLAN, allocate one deep-dive analysis agent per target.

Plan the split before launching anything.

Phase 2 — Discovery / Draft 1

The modes diverge here.

Phase 2 — AUDIT mode

Launch parallel agents, one per facet group. Each agent scans its facet across the scoped paths and returns candidate opportunities with evidence.

Default facet split (7) — run all that apply

Drop any facet that doesn't apply (e.g. drop Data layer for a static site).

Extension facets — add only when relevant

Per-agent prompt pattern (AUDIT)

You are scanning <facet> in the project at <path>.
Project conventions are summarized at <working note path>.

For every refactor opportunity you find, return:
- Short title (≤80 chars)
- One-paragraph summary
- Evidence: 3–5 file:line references (real, verifiable)
- Impact dimensions affected: maintainability / reliability / speed / accuracy / security (rate each H/M/L/—)
- Estimated effort: S / M / L / XL  (S=<1 dev-day, M=1–3, L=3–10, XL=>10 — XL must be split before PLAN)
- Estimated risk: S / M / L
- Why now (or: why later) — one sentence

Constraints:
- Do not propose feature changes, only refactors (same behavior, better internals).
- Do not propose anything without evidence.
- Do not duplicate findings already covered by a more general one — note overlaps.
- Be honest about effort and risk; underestimating burns the executor.

Save findings to <agent-specific path>.

Effort/risk anchors

Effort:
  S  = <1 dev-day
  M  = 1–3 dev-days
  L  = 3–10 dev-days
  XL = >10 dev-days  — MUST be split into smaller items before any PLAN runs

Risk:
  S = isolated; rollback is `git revert`; no shipped state affected
  M = touches multiple callsites or shared abstractions; rollback is straightforward
  L = touches data, migrations, or cross-service contracts; rollback requires explicit plan

Assemble Draft 1 (AUDIT)

1. Cluster duplicates. Two facets often surface the same root issue. Merge with a unified title; keep evidence from all angles.
2. Compute priority score (advisory, not verdict):

   priority = (impact_sum × confidence) / (effort × risk)
   where impact_sum = sum of dimension ratings (H=3, M=2, L=1, —=0)
         confidence = 1.0 if evidence is strong, 0.7 if partial, 0.4 if weak
         effort     = S=1, M=2, L=4, XL=8
         risk       = S=1, M=2, L=4
   

3. Order top-to-bottom by score, then editorially re-order if judgment disagrees. When the agent overrides the score-based order, it must add a one-sentence rationale next to the moved item:

   "Moved REFACTOR-005 above REFACTOR-003 despite lower formula score: blocks 4 other opportunities by clearing a shared abstraction." Reviewers can challenge both the formula inputs and any overrides.

4. One paragraph per opportunity in AUDIT — not an execution spec. The user's next step is to pick top 1–3 for PLAN mode.

Phase 2 — PLAN mode

For each target (1–3), launch one deep-dive analysis agent.

Per-agent prompt pattern (PLAN — discovery half)

You are deeply analyzing the refactor target: <target name + scope>.
Project at <path>. Conventions at <working note path>.

Map the target's full topology:
1. Exact files / modules / functions / endpoints / tables in scope
2. Every callsite of the target's public surface (file:line refs)
3. Every test that exercises the target (file:line refs)
4. Every external dependency the target uses
5. Behavior contract (what does this currently do, observably)
6. Test commands that exercise the target
7. Any manual / browser-driven behaviors that matter
8. Performance characteristics (if measurable)
9. Documented or undocumented edge cases
10. Anything currently broken in this target (separate from the refactor — flag, don't fix)

Save topology to <topology path>. Keep file:line refs precise.

Draft 1 production (PLAN)

After topology agents return, produce one PLAN spec per target, following the PLAN Report Schema below. Drafting can be parallelized (one drafting agent per target) once topology is in.

Phase 3 — Dual Review

Two reviewers in parallel. Different perspectives. Reviewer A is the same in both modes; Reviewer B changes by mode because the right "different perspective" genuinely differs.

Reviewer A — The Skeptic (both modes)

Reads Draft 1, challenges the claims.

Looks for:

• Opportunities (AUDIT) or proposals (PLAN) not backed by evidence
• Overstated impact — "this is critical" when it's a 50-line module nobody touches
• Understated risk — refactors that look small but ripple across many callsites
• Effort estimates that are clearly wrong in either direction (especially: any XL hiding as L)
• Hidden behavior changes sneaking into a "refactor" — flag any proposed change that would alter observable output
• Missing alternatives — was a smaller / less invasive option considered?
• Conventions drift — does the proposed end-state align with the project's own guidelines?

Reviewer B — by mode

AUDIT → Reviewer B = The Prioritizer

Asks: is this ordering defensible?

Looks for:

• Items mis-ranked by the formula (effort or risk under/overstated)
• Editorial re-orderings that lack rationale
• Top items that are easy but low impact crowding out higher leverage items
• Bottom items that should not even be on the list (out of scope, behavior change in disguise, already mitigated)
• Themes / cross-cuts that should have been called out (e.g. five separate items all rooted in the same broken abstraction)
• Confidence ratings that don't match the strength of cited evidence

PLAN → Reviewer B = The Executor's Eye

Asks: can a developer actually run this end-to-end?

Looks for:

• Baseline capture is concrete enough to actually run (commands present, outputs defined, save locations specified)
• Refactor steps are concrete enough to follow without re-thinking the design
• Callsite hunt list is complete (every place that touches the target is enumerated)
• Post-refactor parity check is a real comparison, not "the tests pass"
• Rollback plan is real (not "git revert" — what if the migration is partially applied?)
• HTML artifact template is producible with the data the executor will have
• Drift log discipline is preserved — no hidden behavior changes smuggled in via §9.6

Required Reviewer Output Format

Both reviewers produce reports in this shape:

# Review Report — <Reviewer A | Reviewer B>

## CRITICAL — Will produce a wrong / unsafe / unexecutable report
1. <short title> — <where, what, how to fix>

## HIGH — Will produce confusion or rework
...

## MEDIUM / LOW
...

## Specific challenges
- Claim X (in section Y) is not supported by evidence — needs <citation>
- Effort estimate Z is undercounted because ...
- ...

## Priority Action Items (P0 / P1 / P2)

Vague feedback ("strengthen the verifiability section") is rejected and re-requested. The reviewer must say what line, what fact is missing, what concrete change.

Phase 4 — Draft 2

Synthesize ALL findings from both reviewers into a unified fix list. Then launch parallel rewrite agents (typically same split as Phase 2) with explicit per-section fix instructions.

Draft 2 agent prompt pattern

You are rewriting <section / target> for Draft 2. Read the current draft at <path>.
Apply ALL of the following fixes:

1. <Specific fix — where, what, why> — e.g., "REFACTOR-007's evidence cites no file:line; add 3+ refs from the topology in <topology path>"
2. ...

Save the rewritten section to <new path>.

Fixes must be specific and copy-pasteable. "Strengthen security" is rejected. "Add input validation step at line N of REFACTOR-003's refactor instructions, using the project's zod schema convention from src/lib/validation/" is accepted.

Cross-cutting consistency to reconcile in Draft 2

• Test commands — one canonical command per test tier; same string used in baseline and post-check
• Path conventions — match the project's actual layout, not generic guesses
• Naming of new code — match project's existing naming
• Validation / error model — must match what the project already uses
• Logging / observability — match the project's existing logger and shape
• Environment / config — match the project's single source of truth for env vars
• HTML artifact location & filename pattern — pick once, reference everywhere

Phase 5 — Final Review

Single reviewer, final pass. Goal: catch what the dual review missed.

Specifically check:

• Did Draft 2 fixes actually land?
• Did any fixes introduce new contradictions?
• Are all evidence citations real (file:line) and accurate?
• Is the priority ordering (AUDIT) or per-target ordering (PLAN) defensible?
• Are all test commands runnable as written?
• (PLAN) Does each target's pre-baseline have a 1:1 partner in the post-comparison?
• (PLAN) Is the HTML artifact template self-contained and producible?
• (PLAN) If §7.6 (Known pre-existing failures) has entries, are they justified?
• (PLAN) If §9.6 (Drift log) is anticipated, are the four discipline rules visible?
• Does each milestone / step have a testable exit criterion?
• Beyond this checklist — last chance to catch anything that would make the report stronger.

Output: a concise list of remaining issues. If the doc is solid, the reviewer says so explicitly.

Phase 6 — Draft 3

Single agent applies the final review's remaining fixes. Usually small. After this pass:

• Mark the report Final
• Stamp date and project name
• Save to working location

Phase 7 — Delivery

Three things must happen, adapted to the host environment:

1. Render the report as a shareable HTML artifact — the AUDIT or PLAN goes through the project's artifact-display mechanism so the user sees and can share it.
2. Save a portable copy — a markdown file at a stable path; if the host supports public links, produce one.
3. Update working memory — drafts produced, key decisions, what mode was used, what targets if PLAN.

For PLAN mode, the deliverable to the executor developer must include the HTML artifact template they'll fill in — see "Executor's HTML Artifact Template" below.

AUDIT Report Schema

Header:

• Mode: AUDIT
• Project:
• Scope: <paths or "full project">
• Date:
• Methodology: Multi-round, multi-agent (Code Refactor skill)

Sections:

Executive Summary

3–5 sentences. Overall health, top 1–3 themes, suggested next step (typically: "run PLAN on REFACTOR-001..003").

Scope & Method

• Paths reviewed
• Facets reviewed (note any default facets dropped, any extension facets added)
• Conventions referenced
• Tools / agents used

Priority Scoring

Show the formula. Show the H/M/L scale and effort/risk anchors. Note that the score is advisory — the agent may editorially re-order items, with one-line rationale per override.

Opportunities (ranked)

For each opportunity, in priority order:

### REFACTOR-NNN — <Title>

**Scope:** <files / modules / endpoints>

<One paragraph: what's wrong, what's the better state.>

**Evidence:**
- `path/to/file.ts:47–93` — <what's there>
- `path/to/other.ts:12` — <what's there>
- ...

**Impact:**
- Maintainability: H / M / L / —
- Reliability: H / M / L / —
- Speed: H / M / L / —
- Accuracy: H / M / L / —
- Security: H / M / L / —

**Effort:** S / M / L / XL
**Risk:** S / M / L
**Confidence:** 1.0 / 0.7 / 0.4
**Priority Score:** <number>
**Editorial override (if any):** <one sentence rationale, omit otherwise>

**Why now / why later:** <one sentence>

No execution detail in AUDIT — that's PLAN's job.

Themes & Cross-Cuts

If multiple opportunities share a root cause (e.g. "validation drift across endpoints"), call it out as a theme. Helps the user pick PLAN targets that have leverage.

Recommended PLAN Targets

Top 1–3 the user should consider for deep-dive. State why each made the cut.

PLAN Report Schema (per target)

For each of 1–3 targets, the PLAN contains exactly the sections below. Numbering is per-target (REFACTOR-007 → P-007.1 through P-007.12).

P-NNN.1 Title & ID

• ID: matches the AUDIT ID if applicable, else a fresh REFACTOR-NNN
• Title
• Date:
• Project:
• Branch convention: <e.g. refactor/<slug> per project guidelines>

P-NNN.2 Scope

Exact files / modules / endpoints / tables / migration files in scope. Anything not listed is out of scope.

P-NNN.3 Current State

• What's there now
• Code excerpts (≤30 lines each, file:line cited)
• Callsite map (every place that uses the target's public surface)
• Tests that currently exercise it (file:line)
• Behavior contract — what the target observably does today, in plain prose

P-NNN.4 Problem Statement

Why this needs refactoring. Concrete, evidence-bound. Not "code smell" — what smell, where, why it matters.

P-NNN.5 Refactor Goal

What the better state looks like. Architecturally, in prose. Plus: the non-goals — what this PLAN deliberately does not change. (Non-goals stop scope creep.)

P-NNN.6 Impact Dimensions

H/M/L per dimension (maintainability, reliability, speed, accuracy, security), with a sentence justifying each rating.

P-NNN.7 Pre-Refactor Verifiability — BASELINE

The executor MUST complete this section in full before touching any code.

7.1 Environment baseline

Exact commands; expected: pass.

• <typecheck command> — expect: 0 errors
• <lint command> — expect: 0 errors
• <build command> — expect: success
• <full test command> — expect: all pass

If any of the above does not currently pass on the base branch, the executor STOPS and chooses one of:

• (a) Fix the failing check first, in a separate PR. Refactor PR rebases on top once base is green. Default option.
• (b) Scope the failing check out of the parity guarantee — see §7.6.

7.2 Targeted test capture

• List of tests that exercise the target (commands that run them in isolation)
• Save outputs verbatim to <baseline path>/tests-pre.txt

7.3 Behavioral baseline (manual smoke tests)

For each user-visible behavior the target affects:

• Step-by-step (click here, enter this, hit submit)
• Expected output (exact strings, screenshot if applicable)
• Save observations to <baseline path>/manual-pre.md

7.4 Performance baseline (if relevant)

• Exact command(s) to capture timing / memory / bundle size
• Capture 3 runs, record median
• Save raw numbers to <baseline path>/perf-pre.txt

7.5 Baseline checkpoint

Commit baseline files to a fresh branch (refactor/<slug>-baseline) so they're recoverable.

7.6 Known pre-existing failures (escape hatch — use sparingly)

If §7.1's hard STOP rule is invoked under option (b), each excluded check is documented here:

- Check: `<exact command>`
- Failure: <one-line description; link to issue if open>
- Why scoped out: <why fixing-first is impractical for this PLAN>
- Parity impact: <which behaviors this check normally guards; what risk this exclusion creates>
- Reviewer/PM acknowledgment: <name + date>

Rules:

• §7.6 is empty by default. Only entries explicitly invoked under §7.1(b) appear here.
• A PLAN with more than one entry in §7.6 signals the AUDIT mis-scoped this target. Reviewer A flags this in Phase 3.
• Each entry must include reviewer or PM acknowledgment before the PLAN ships to an executor.

The baseline must be concrete and runnable. If the executor has to invent commands, the PLAN failed.

P-NNN.8 Refactor Instructions — THE WORK

Pre-conditions:

• Branch refactor/<slug> created from <base>
• Baseline section 7 fully completed
• All baselines green (or §7.6 entries acknowledged)

Step-by-step changes (numbered):

For each step:

• What changes (which file, which lines)
• Proposed diff or precise edit instruction
• Why this step (one sentence)
• Convention notes (link to project guidelines)
• Gotchas / things to pay attention to

New abstractions / interfaces introduced:

• Name, signature, location
• Why it earns its keep

Migration approach:

• In-place edit, parallel-implementation-with-cutover, or shim-and-deprecate?
• If multi-step, the cutover sequence

Callsite hunt list:

• Every callsite (file:line) that must be updated
• For each: "before / after" delta description

Things to pay attention to:

• Cross-cutting concerns the executor will hit (logging, error model, validation)
• Project conventions to honor

Style & convention notes:

• Direct references to the project's coding standards
• Naming, file org, formatting expectations

P-NNN.9 Post-Refactor Verifiability — PARITY CHECK

The executor MUST complete this section in full before declaring done.

9.1 Re-run environment baseline

• Same commands as 7.1, expect: same results (still all green, except any §7.6-acknowledged checks)

9.2 Re-run targeted tests

• Same commands as 7.2
• Save outputs to <baseline path>/tests-post.txt
• Diff against tests-pre.txt. Differences require explicit justification (e.g. test was renamed in this refactor, or covered under §9.6 drift log) — note in §9.6.

9.3 Re-run manual smoke tests

• Same scripts as 7.3
• Save observations to <baseline path>/manual-post.md
• Behavioral parity checklist:
  • Behavior 1 — same as pre
  • Behavior 2 — same as pre
  • …

9.4 Re-run performance baseline

• Same commands as 7.4 (3 runs, median)
• Save to <baseline path>/perf-post.txt
• Compute delta vs pre. Speed-target refactors must show measurable improvement; non-speed refactors must show no regression beyond <X%> noise floor (define X% per project).

9.5 New tests added (optional but encouraged)

If the refactor surfaces a previously-untested edge case, add a test for it. Note location.

9.6 Drift log — intentional behavior changes (use sparingly, strict rules)

If, during the refactor, the executor finds a pre-existing bug whose fix is unavoidable in-flight, document each occurrence here. Empty by default. Each entry must satisfy ALL four rules:

Drift entry:
  Location:        <file:line of the changed behavior>
  Pre-existing bug: <commit / issue / code reference proving the bug existed before this refactor>
  Diff justification: <pre-output → post-output; why it differs>
  New test added:   <path:line of test that locks in the corrected behavior>

Four discipline rules:

1. Pre-existing only. Drift may only correct behavior that was already wrong before the refactor began. Net-new functionality is not drift.
2. Documented in §9.6. Every diff vs. baseline that is not byte-identical must be either covered here or in a renamed-test note in §9.2.
3. Test required. Every drift entry has a corresponding new test that locks the corrected behavior.
4. Size cap. If the drift would change behavior in more than ~3 places, the executor STOPS. The work splits: refactor PR (no drift) ships first; bug fixes ship as a separate PR. The PLAN itself does not absorb the fix scope.

If the executor cannot satisfy all four rules for a behavioral diff, the refactor has changed behavior — STOP and treat as a feature/fix, not a refactor.

P-NNN.10 Acceptance Criteria

All must be true to declare done:

• All env baseline checks green (or §7.6 entries acknowledged)
• All targeted tests pass
• tests-pre.txt ≡ tests-post.txt (or differences justified in §9.2 / §9.6)
• All manual smoke tests pass with same observable behavior (modulo §9.6 drift entries)
• Perf check passes (improvement for speed targets; no regression otherwise)
• No new lint / typecheck errors
• Project conventions honored (per project guidelines)
• §9.6 drift log: ≤ ~3 entries, all four rules satisfied, all backed by new tests
• HTML artifact (§12) produced and saved
• PR opened with link to artifact

P-NNN.11 Rollback Plan

• If parity fails partway: branch is disposable; revert to baseline branch.
• If a partial migration is shipped: explicit unwind procedure (DB migration down-migration, feature-flag flip, cutover reversal).
• If post-merge regression discovered: the revert PR command, plus any data fixes needed.

The rollback must be specific. "Git revert" is not a plan unless the change is purely code with no migrations / no shipped state.

P-NNN.12 Final Report — HTML Artifact

The executor produces an HTML artifact summarizing the run. Use the Executor's HTML Artifact Template (below) as the starting point.

Required artifact contents:

• Header: target name, date, executor, branch, PR link
• Pre-baseline summary (commands + key outputs)
• Changes summary (files changed, LOC delta, key abstractions added)
• Post-baseline summary (commands + key outputs)
• Diff vs baseline (tests, behavior, perf)
• §7.6 entries (if any) — acknowledged exclusions
• §9.6 drift log (if any) — entries with proof
• Acceptance checklist (filled in)
• Verdict: PASS or FAIL (with reason)

Save to <artifact path> and surface via the project's artifact-display mechanism.

Executor's HTML Artifact Template

A self-contained HTML document the executor fills in at the end of a PLAN. The PLAN must hand this template (or a pointer to it) to the executor.

Structure (in plain HTML/CSS, no external deps):

- <header>: Project, Target ID, Title, Date, Executor, Branch, PR
- <section id="verdict">: Big PASS / FAIL banner, one-sentence summary
- <section id="pre-baseline">:
    - Env checks: command + result table
    - Test results: pre snapshot
    - Manual checks: pre snapshot
    - Perf numbers: pre snapshot (3 runs + median)
    - §7.6 acknowledged exclusions (if any)
- <section id="changes">:
    - Files changed (table)
    - LOC delta
    - New abstractions introduced
    - Notable design decisions
- <section id="post-baseline">:
    - Env checks: command + result table
    - Test results: post snapshot
    - Manual checks: post snapshot
    - Perf numbers: post snapshot (3 runs + median)
- <section id="diff">:
    - Test output diff (pre vs post) — must be empty or fully justified
    - Behavior parity checklist (filled)
    - Perf delta table
    - §9.6 drift log entries (if any)
- <section id="acceptance">: checklist with checks filled in
- <section id="rollback">: link to rollback plan, plus disposal instructions if shipped

Style: simple, monospace for command/output blocks, clean tables, no external CDN. Mobile-responsive. The template itself is project-agnostic; the PLAN customizes contents.

Anti-Patterns to Avoid

Quality Markers

A finished report should pass:

• Mode is unambiguous; outputs match the mode's schema
• Every claim has evidence (file:line) — none are hand-waved
• (AUDIT) Priority ordering is defensible; formula is visible; editorial overrides are justified
• (AUDIT) No XL items remain — they've been split
• (PLAN) Pre and post baselines are 1:1 mirrors — every check has a partner
• (PLAN) §7.6 has 0 or 1 entries (more = AUDIT was wrong); each entry is acknowledged
• (PLAN) §9.6 drift log is empty by default; if populated, all four rules satisfied per entry
• (PLAN) Every step in §8 is concrete enough to execute without re-thinking
• (PLAN) Callsite hunt list is exhaustive (the executor doesn't have to grep for more)
• (PLAN) Acceptance criteria are testable, not aspirational
• (PLAN) Rollback plan is real (not just "git revert")
• (PLAN) HTML artifact template is self-contained and producible
• Project conventions are honored end-to-end
• No surviving "TBD", "TODO", or "we'll figure this out later"
• The receiving developer can run the loop without coming back with clarifying questions

Reference Workflow Checklist

Phase 1 — Scoping & Context Load
  [ ] Mode locked: AUDIT or PLAN
  [ ] Targets locked (if PLAN): 1–3 specific items
  [ ] Scope locked: paths / subsystems
  [ ] Project conventions loaded (CLAUDE.md, dev guidelines, PRD)
  [ ] Test/build/lint commands captured
  [ ] Discovery split planned (default 7 facets for AUDIT, drop/add as fits)

Phase 2 — Discovery / Draft 1
  [ ] (AUDIT) Default facets selected; extension facets added if relevant
  [ ] (AUDIT) Facet agents launched in parallel
  [ ] (PLAN) One topology agent per target; launched in parallel
  [ ] All agents returned
  [ ] Findings clustered (AUDIT) / topology assembled (PLAN)
  [ ] (AUDIT) XL items split before they reach the ranked list
  [ ] Draft 1 written per the appropriate schema

Phase 3 — Dual Review
  [ ] Skeptic reviewer launched
  [ ] Mode-appropriate Reviewer B launched (Prioritizer for AUDIT, Executor's Eye for PLAN)
  [ ] Both returned with required output format
  [ ] Findings unified into a single fix list

Phase 4 — Draft 2
  [ ] Per-section fix instructions specific & copy-pasteable
  [ ] Rewrite agents launched in parallel
  [ ] Cross-cutting consistency reconciled
  [ ] Draft 2 assembled

Phase 5 — Final Review
  [ ] Single reviewer pass
  [ ] Remaining issues listed (concise)

Phase 6 — Draft 3
  [ ] Final fixes applied
  [ ] Marked Final + dated

Phase 7 — Delivery
  [ ] Rendered as HTML artifact via host's artifact mechanism
  [ ] Markdown copy saved to stable path
  [ ] Working memory updated
  [ ] Handed off to user with clear pointers (and PLAN handoff instructions if applicable)

When to Update This Skill

Update when:

• A new facet emerges that AUDIT consistently misses (add to the default 7 or extension menu)
• A new category of "executor blocker" appears in PLAN reviews (add to Executor's Eye checklist)
• A new project type needs a different verifiability shape (e.g. mobile app smoke testing)
• A pattern emerges that should become a default

Do not update for:

• Project-specific quirks (those go in the report itself, not the skill)
• One-time formatting preferences
• Tech-stack-specific opinions (those belong in code-conventions skills)