About review-implementation
An internal classification skill used by review and audit workflows to assess synthesized findings. It checks coverage of applicable review axes, classifies each finding, and routes it without changing code.
Key features
- 01Checks review-axis coverage
- 02Classifies synthesized findings
- 03Routes findings by current-unit contract
- 04Uses a shared classification rubric
- 05Does not edit code
Use cases
- 01Classifying findings from review-change
- 02Routing findings in audit-pr reviews
- 03Applying a common rubric to product audits
Install
Copy the command and run it in a terminal.
npx skills add https:/ / github.com/ gtrabanco/ agentic-workflow --skill review-implementationAfter installing, type/review-implementationin Claude Code to confirm it loaded.
Bundled filesreferences/CLASSIFY.mdreferences/FIND.md
Details
From the project READMEReview Implementation (internal scope/classification engine)
The classification engine the review/audit skills compose: it consumes the
synthesized findings table (the fused output of the applicable per-axis
passes), verifies every applicable axis is represented, and returns the
classified decision table — then stops. Never refactors or edits code. It owns
the scope/axis-coverage contract and the classification rubric (the
current-unit contract + routing) that review-change, audit-pr, and
product-audit reference instead of restating.
It does not scan the diff: every finding concern has exactly one owning
pass (see the axis ownership map) — the per-axis passes
(review-code, review-security, review-verify, review-perf,
design/a11y/brand/SEO) find, and this engine classifies. No broad findings
scan here.
When to use
- Invoked by
review-change(the user-facing review entry) as its classification engine, over the fused findings table. - The audit skills reference its rubric and coverage contract.
Scope
The caller's scope statement (the branch diff vs. the default branch, or the passed path/glob) is authoritative; the synthesized table was gathered over it. State the scope at the top of the classified report.
Step 0 — Discover the project (always first)
Per the agent guide's Workflow conventions + documentation map, read
what THIS skill needs: the architecture/layering rules, the testing philosophy,
and any runtime/platform, security, money, i18n/SEO/a11y and bundle rules. Pull
the project's specific risk axes from its guardrail skills where present. The
FIND.md axis map is the default; the project's docs refine which axes are
applicable.
Step 1 — Verify axis coverage (the synthesized table)
For the declared scope, confirm every applicable axis is represented in the
synthesized findings table — one finding owner per axis, per the FIND.md
map: an axis the change touches that the table says nothing about is a
missing-axis finding (axis coverage), not a silent pass. Overlapping
signals from different passes on the same defect collapse into one row during
synthesis — the table must contain neither duplicates nor gaps. State which
axes were applicable and confirm each appears.
Step 2 — Classify (the current-unit contract)
Read Classify and route and classify every row of the
synthesized table without reopening source files: ignore first (the claim),
then the current-unit contract (fix-now / replan-in-unit / decision-required
for in-scope work), then proposal for genuinely independent future
capabilities. One pass — no per-pass or per-reviewer classification.
Context budget
The input is the synthesized table, not the diff. Read at most 10 non-diff files in full for surrounding context (callers, contracts, SPEC); targeted reads (≤ 50 lines of a named range) and grep/glob results don't count. Record each classification as its table row immediately and drop raw file content.
Guardrails
- Findings + table only. Never refactor or edit code in this skill.
- One classifier. Classification happens HERE, once, over the fused table — never per-reviewer, never re-litigated in the per-axis passes.
- Honor the dead-code exception — staged/planned code is not dead code.
- Don't inflate severity; separate "correctness/security" from "taste".
- Don't deflate either: current-unit work is never
postpone/tradeoff/wontfix/disputedand never a new issue — size routes toreplan-in-unit, not to a downgrade (current-unit contract inCLASSIFY.md). - Otherwise per the project's Workflow conventions (docs-language,
evidence): cite
file:line, mark uncertainties verify.
Relationship to other skills
- Classification engine of
review-change— the user-facing review skill runs the applicable per-axis passes (the finders), fuses their tables, then composes this engine to classify.audit-prandproduct-auditreuse this rubric. - Sits in Stage 4 of the feature workflow (verification & review).
fix-nowfolds into the current unit;replan-in-unitrunsnode scripts/unit-route.mjs <unit>, whoseroute: replanline names the planner that appends user-confirmed phases beforeexecute-phase;decision-requiredblocks for the user; independent work becomes proposals the user routes totriage-issue(D3).
Done when
- A synthesized table consumed, axis coverage verified (no applicable axis missing, no duplicate rows), every finding classified with reasoning and routed — and no code changed.