design-review

Overview

design-review performs an adversarial pre-approval review of a design document — a spec, design doc, RFC, ADR, or implementation plan. It hunts for the recurring gaps that otherwise surface late (or in production): missing bootstrap/ownership, misleading names, unconsidered scale, unstated assumptions, claims that don't match the code, missing failure handling, unexamined security surface, unjustified complexity, and incompleteness. Critically, it verifies claims against the actual codebase rather than trusting the prose, and it cites file:line evidence so every finding is grounded. It produces a findings report and a verdict; it never edits the document and never approves — that stays with the human reviewer and the author. Use it as the gate between writing a design and signing off on it.

When to activate

• ✅ A spec / design doc / RFC / ADR / implementation plan is about to be approved and you want gaps found first.
• ✅ A document author wants a critical pass before presenting their design for sign-off.
• ✅ An automated workflow needs a consistent pre-approval review step over design documents.

Do NOT activate when:

• The artifact is code / a diff — use a code-review skill (requesting-code-review / code-review); this skill reviews design documents, not implementations.
• You are authoring the document — use a content/template skill to draft it; design-review only reviews an existing draft.
• The document is the doc-library technical-design artifact — identified authoritatively by a template: technical-design frontmatter (a # Technical Design: heading is a fallback signal only when frontmatter is absent; a generic doc that merely titles itself "Technical Design" without the stamp stays here) — use reviewing-technical-design, its dedicated reviewer that asserts the TDD-specific implementability bar (requirement trace, reference-not-duplicate, observability, testing, rollout, delta-scoped amend). design-review still gates generic design docs, RFCs, ADRs, specs, and plans — only the doc-library TDD artifact is carved out.
• The document is the doc-library architecture-doc artifact (a whole-system architecture document + its linked ADR files) — identified authoritatively by a template: architecture-doc frontmatter (a # Architecture: heading is a fallback signal only when frontmatter is absent; a generic doc that merely titles itself "Architecture" without the stamp stays here) — use reviewing-architecture-doc, its dedicated reviewer that asserts the architecture-quality bar (boundary+concerns, structure+altitude, diagram-sync, the ADR mechanism, NFR-realization, cross-cutting, ASR coverage, delta-scoped amend). Boundary (load-bearing): this carves out only the architecture-doc-library artifact and the ADR files it links; a standalone ad-hoc ADR or RFC (not part of an architecture-doc artifact) stays with design-review.
• The document is the doc-library data-model artifact (a persistence/domain model — entities, keys, relationships, indexes, normalization, lifecycle) — identified authoritatively by a template: data-model frontmatter (a # Data Model heading is a fallback signal only when frontmatter is absent; a generic doc that merely titles itself "Data Model" without the stamp stays here) — use reviewing-data-model, its dedicated reviewer that asserts the integrity + queryability bar (typed+keyed entities, cardinality + referential rule, access-pattern-justified indexes, normalization + paradigm choice, lifecycle/migration, diagram⇄tables sync, one-directional-vs-the-api-spec, delta-scoped amend). The api-spec / feature-spec the model references stay with their own reviewers; only the data-model artifact is carved out.
• The document is the doc-library api-spec artifact (an engineering API wire contract — operations, request/response schemas, auth, the error model, pagination, versioning, examples) — identified authoritatively by a template: api-spec frontmatter (a # API Specification heading is a fallback signal only when frontmatter is absent; a generic doc that merely titles itself "API Specification" without the stamp stays here) — use reviewing-api-spec, its dedicated reviewer that asserts the contract-completeness bar (every operation typed both sides + status codes, the complete error model, shared types referencing the data-model, per-operation authorization, pagination + tie-breaker, examples matching the schemas, one-directional-vs-the-api-reference, delta-scoped amend). Style-agnostic (REST/GraphQL/gRPC). The data-model / feature-spec it references + the downstream api-reference stay with their own reviewers; only the api-spec artifact is carved out.
• The text is a one-line note or non-structured prose with no design to evaluate.

Workflow

Step 1: Receive inputs

• Required: a path to the document to review.
• Optional: explicit codebase pointers (dirs/modules the design touches). If omitted, infer the modules to verify from the document's own references (file paths, command/module names, links it cites).
• Optional: an output preference — inline (default) or write a review.md beside the document.

Step 2: Read the document fully

Read the whole document before judging any part. Note its claims about existing behavior, its named commands/files/verbs, its inputs, its scale assumptions, and any code paths it references — these feed Steps 3–4.

Step 3: Gap-hunt by category

Walk the document against this rubric. The list is extensible — add categories as new gap types surface; never drop the core ones.

Plan lens — when the document under review is an implementation/project plan, additionally check (the 9 categories above still apply):

Step 4: Verify claims against the code (mandatory)

For any finding that asserts something about existing behavior, paths, or interfaces, read/grep the referenced module and confirm it — then cite file:line. This is the difference between a grounded review and a guess.

• A finding asserting an inconsistency must quote/point at the real code (path:line).
• A gap you suspect but cannot confirm is marked unverified explicitly — never asserted as fact (this enforces "no fabrication").
• Greenfield: if the document describes brand-new behavior with no existing code to check, the consistency category is N/A — absence of code to verify is never itself a blocker.
• Bound the work: verify only the modules the document references; do not scan the whole tree.

Step 5: Assemble findings

Each finding:

- [severity] <category> — <location in the doc>
  Gap: <what is missing/wrong/unstated>
  Fix: <concrete suggested resolution>
  Evidence: <path:line> | unverified

Severity: blocker (must resolve before approval — wrong/missing in a way that breaks the design or its premises) · important (should resolve — real gap, not fatal) · minor (nice to fix).

Step 6: Verdict

End with one line: ready-for-approval (no blockers) or has-blockers (list the blockers first, then important, then minor). The verdict is a recommendation to the human — not an approval.

Step 7 (optional): Independent pass for large / high-risk docs

For a large or high-stakes document, dispatch a fresh reviewer with no prior context (a subagent) given the same rubric, then merge its findings (de-duplicated). If subagent dispatch is unavailable, fall back to a second inline pass and say so. This mirrors the value of a fresh set of eyes; it is optional, not required.

Step 8: Output

• Inline by default: return the findings + verdict to the caller.
• Opt-in: if requested, write a review.md beside the document.
• Never edit the document. Never approve. Recommend; the author applies, the human decides.

Rules

Hard rules (never violate):

• Verify before asserting (no fabrication). Any claim about existing behavior is confirmed against the code with a file:line citation, or it is labelled unverified. Never invent a gap.
• Review-only. Never edit the document; never approve or reject it. Output is findings + a verdict recommendation; the human owns the decision and the author owns the fix.
• Greenfield is not a gap. Absence of code to verify against is never, by itself, a blocker.
• Bound verification to the modules the document references — do not scan the whole repository.
• Severity honesty. Reserve blocker for things that genuinely must be fixed before approval; do not inflate.

Preferences (override-able):

• The category rubric is extensible — add categories as new recurring gap types emerge.
• Surface blockers first; keep the report scannable; collapse minors if numerous.
• Prefer one concrete suggested fix per finding over open-ended commentary.

Gotchas

• Trusting the prose. A spec can confidently describe behavior the code doesn't have. The whole value is Step 4 — verify against the code; a review that skips it is just opinion.
• Greenfield mis-flagged. On a brand-new-system doc there is nothing to verify against; flagging "couldn't verify X" as a problem is wrong — mark consistency N/A.
• Whole-tree scan. Trying to verify against the entire codebase is slow and noisy; bound it to referenced modules.
• Over-flagging. Listing every nitpick as a blocker buries the real ones. Severity discipline keeps the verdict meaningful.
• Detail mistaken for completeness. Lots of detail is not the same as a complete design; conversely, terse-but-complete is fine. Judge whether the decisions are present, not the word count.

Anti-patterns

• "It probably doesn't match the code" without checking. Fabricated or unverified gaps stated as fact — forbidden; verify or mark unverified.
• "I'll just fix the doc while I'm here." Editing the document is out of scope; review-only.
• "Looks good to me." Rubber-stamping — skimming and approving without running the rubric or verifying claims.
• "Let me approve it." Deciding the outcome; the skill recommends, the human approves.
• "Let me grep the entire repo to be safe." Unbounded verification; stay within referenced modules.

Output

A findings report (inline by default; optional review.md): zero or more findings in the Step-5 shape, ordered by severity, followed by a one-line ready-for-approval | has-blockers verdict. The artifact is a recommendation consumed by the document's author and the human approver at the approval gate — never an approval, never an edited document.

Related

• requesting-code-review / code-review — the analogous capability for code/diffs; design-review is the design-document counterpart that runs before code exists.
• A content/template skill — authors the document that this skill reviews (authoring vs reviewing are distinct).
• reviewing-technical-design — the dedicated reviewer for the doc-library technical-design artifact (a TDD), which is carved out of this skill's scope. Route a template: technical-design document there; design-review keeps generic design docs, RFCs, ADRs, specs, and plans.
• reviewing-architecture-doc — the dedicated reviewer for the doc-library architecture-doc artifact (a whole-system architecture document + its linked ADR files), also carved out of this skill's scope. Route a template: architecture-doc document there; a standalone ad-hoc ADR/RFC (not part of an architecture-doc artifact) stays with design-review.
• reviewing-data-model — the dedicated reviewer for the doc-library data-model artifact (a persistence/domain model), also carved out of this skill's scope. Route a template: data-model document there; design-review keeps generic design docs, RFCs, ADRs, specs, and plans.
• reviewing-api-spec — the dedicated reviewer for the doc-library api-spec artifact (an engineering API wire contract), also carved out of this skill's scope. Route a template: api-spec document there; design-review keeps generic design docs, RFCs, ADRs, specs, and plans.
• Fits the spec → plan → implement discipline as the pre-approval gate over the spec (and plan).

Progressive disclosure

• references/sources.md — research provenance for the rubric. Load only if auditing where the categories came from.

This skill ships no scripts/ or assets/; it runs via Read / Grep / Glob and (optionally) a subagent dispatch.

Body budget

• description ≤ 1,024 chars (agentskills.io cap); combined description + when_to_use truncated at 1,536 chars in the listing.
• Body ≤ ~500 lines / 5,000 tokens.

Changelog

• 2.4.0 (2026-06-15) — scope carve-out: the doc-library api-spec artifact now routes to its dedicated reviewer reviewing-api-spec; the data-model/feature-spec it references + the downstream api-reference stay with their own reviewers. Additive only — the 9 gap categories and the plan lens are unchanged. (The fourth and last authoring-only carve-out — every doc-library type now has a dedicated reviewing twin.)
• 2.3.0 (2026-06-15) — scope carve-out: the doc-library data-model artifact now routes to its dedicated reviewer reviewing-data-model; the api-spec/feature-spec it references stay with their own reviewers. Additive only — the 9 gap categories and the plan lens are unchanged.
• 2.2.0 (2026-06-14) — scope carve-out: the doc-library architecture-doc artifact (+ its linked ADR files) now routes to its dedicated reviewer reviewing-architecture-doc; standalone ad-hoc ADRs/RFCs stay here. Additive only — the 9 gap categories and the plan lens are unchanged.
• 2.1.0 (2026-06-14) — scope carve-out: the doc-library technical-design artifact now routes to its dedicated reviewer reviewing-technical-design; design-review keeps generic design docs, RFCs, ADRs, specs, and plans. Additive only — the 9 gap categories and the plan lens are unchanged.
• 2.0.0 (2026-06-01) — initial reviewed release.