Review design doc
Skill gordon1210/agent-goodies/plugins/idea-workbench/skills/review-design-doc
Portable agent plugins for Codex/ChatGPT and Claude Code/Desktop.
npx -y skills add gordon1210/agent-goodies --skill review-design-docAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 19 days oldThe repository was created 19 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Evaluate an existing product or technical design document for comprehension, consistency, risk coverage, and implementation readiness. Use before approval or implementation planning, after substantial design changes, or when a team needs an independent critique that exposes missing decisions, contradictions, unsupported assumptions, unsafe failure modes, and verification gaps without rewriting the source by default.
SKILL.md
4.9 KB, as published. Nobody here has run it
Review a Design Doc
Judge whether a capable reader can safely act on the document as written. Treat the review as read-only unless the user explicitly asks to revise the source.
Preserve a fresh-reader view
- Read the design document before reconstructing intent from surrounding conversation.
- Summarize the problem, chosen direction, boundaries, and validation approach using only what the document communicates.
- Record ambiguities rather than silently resolving them from your own knowledge.
- Then read linked briefs, research, decision records, and relevant project material to verify alignment and factual claims.
- Verify the stated project context. Do not treat missing repository access as proof that the work is greenfield.
- Label any inference that is not stated or evidenced by the source.
Set the review depth
Classify the design as low, medium, or high risk using reversibility, blast radius, novelty, data sensitivity, migration difficulty, external dependencies, and operational burden. Review proportionally, but always examine the core decision, boundaries, behavior, uncertainty, and validation.
Read review-method.md before performing the detailed review. Use its coverage lenses, severity definitions, and verdict rules.
Review in two passes
Test comprehension
Act as a capable implementer who was absent from prior discussions. Determine whether the document makes the following clear without guesswork:
- The problem, desired outcome, scope, and non-goals.
- The chosen direction, rejected alternatives, and accepted trade-offs.
- The important flows, boundaries, contracts, state changes, and ownership.
- The assumptions and open decisions that remain.
- The evidence that will demonstrate success and safe delivery.
Challenge the design
Try to make the proposal fail. Walk realistic success, failure, recovery, misuse, migration, concurrency, and boundary scenarios through it. Look for contradictions across sections and linked artifacts, hidden dependencies, unsupported claims, irreversible steps, unowned operational work, and validation that proves only the happy path.
For greenfield work, challenge undecided foundations and unsupported bootstrap assumptions. For an existing system, challenge integration evidence, compatibility, migration, and preserved behavior. For hybrid work, test the contract, ownership, and failure modes at the new-to-existing seam.
Do not manufacture issues to appear thorough. Prefer a few consequential findings over a catalogue of stylistic preferences.
Report actionable findings
Copy and adapt design-review-template.md. For every finding:
- Assign a stable identifier and severity.
- Point to the relevant section or state that the material is absent.
- State the evidence, not merely a conclusion.
- Explain the credible impact or failure scenario.
- Describe the decision, evidence, or clarification needed to resolve it without prescribing unnecessary implementation detail.
Separate blockers from improvements. Do not downgrade a missing core decision because a likely answer seems obvious, and do not elevate a wording preference into a delivery risk.
Give an explicit verdict
Choose exactly one verdict using the rules in the review method:
- Ready: Safe to proceed to implementation planning.
- Ready with follow-ups: Safe to plan only under the listed, bounded conditions.
- Not ready: Requires design decisions or evidence before planning can proceed responsibly.
State the verdict first, justify it briefly, and list any conditions. If no actionable findings remain, say so and name the residual risks or untested areas instead of implying certainty.
Apply the approval gate after the verdict:
- For Ready, ask the user to approve the reviewed design for implementation planning or request revisions.
- For Ready with follow-ups, ask the user to accept the listed conditions explicitly before treating the design as approved.
- For Not ready, do not ask for planning approval; recommend the smallest design revision or evidence-gathering step that addresses the blocking findings.
Record approval only when the user states it explicitly. A positive verdict, silence, or agreement with one finding is not approval.
Present the review separately from the source document. Stop after the verdict, findings, and approval request: do not edit the design, create an implementation plan, or begin implementation unless the user explicitly asks for that additional work.