Skip to content

rizanw/steno-spec / packages/skills/skills/steno-review

steno-review

Verified

steno-review — zero-history review

Why: a reviewer that shares your conversation agrees with you. Reviewers must see only the packet. A review settles and removes. It never grows the spec on its own, and it ends: one full round, at most one verify round.

Steps

  1. steno review --stage <requirements|design|tasks|drift> → writes reviews/<date>-<stage>-packet.md and prints the round (full or verify), the size and any warning. Note the path; tell the user every warning. If it says nothing changed, go to step 7.
  2. EGM, only on request. If .steno/config.yaml sets reviewer: egm or the user asked for EGM in this request, and EGM is installed: invoke egm-review on that packet file (for drift, egm-drift-report), telling it which stage the file is; for requirements or tasks, add that it is not a design document, so no architecture, alternatives or risks sections are expected. Save its findings as reviews/<date>-<stage>.md. Go to step 5.
  3. Otherwise (the default) check whether this session can start sub-agents that do not inherit the conversation. If it cannot, stop and tell the user a fresh-context review is not possible here; do not review in your own context.
  4. Start one sub-agent per persona in reviewers/ that applies to the stage (each file says which); in a verify round leave out scope. If a sub-agent type named steno-reviewer-<persona> is registered (Claude Code after steno init), use it and give it only the packet path. Otherwise start a plain sub-agent and give it only: the persona file, the packet path, read-only access. Merge duplicates and write reviews/<date>-<stage>.md (append # Round N after the first) in three groups: ## Decide, ## Fix, ## Notes for the next stage, each line - [sev] <id>: <finding> → <change or options>.
  5. Decide first. Ask the user each decide finding, scope cuts before additions, one question at a time with the options and your pick. Never apply a decide on your own, and never turn one into new text before the user says yes. Write each answer and each rejected finding as one line in reviews/<stage>-decisions.md (- <id>: <decision>, <reason>); the next packet shows it to reviewers so they do not reopen it. A cut goes to ## Non-goals with its trigger. A split is not a cut: before deleting anything, run steno new <slug> --title "<Title>" for the part that leaves, move its requirements, glossary terms and open questions into that spec's requirements.md (renumbered from R1; steno lint there), and only then replace them here with one ## Non-goals line naming the new spec. Tell the user the new spec exists and is still at requirements.
  6. Then fix. Drop every fix on an item that was cut. Apply the rest: a fix rewrites or deletes existing text and adds no criterion, section or task. Leave note findings in the findings file; do not edit them into the document. Run steno lint, then list each applied fix to the user in one line.
  7. Gate. The review has passed when no high fix is open and every decide has an answer. Say so, give the size before and after, and offer steno approve <stage>. Offer one verify round (step 1 again) only if step 6 changed the document and a high finding was among the fixes. After the verify round, or when the CLI warns about the round limit, do not start another: list what is still open and let the user approve, edit or ask for steno review --full.
  8. Open questions. If the document still lists entries under ## Open questions, list them numbered and ask the user to pick: Interview (offer first; one question at a time with your recommended answer and its trade-off; after each answer edit the document, delete the question (and the heading with the last one), confirm in one line; "skip" and "stop" are allowed; finish with steno lint), Edit myself (when the user says done, run steno lint and re-list what is open), or Skip (steno approve <stage>; the questions are settled in the next stage). Skip this step for drift.

Personas (reviewers/)

scope · completeness · clarity · code-fit · implementability · ears

receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementationStars: 322kUpdated Claude CodeCodexVerifiedInstall
requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirementsStars: 322kUpdated Claude CodeCodexVerifiedInstall
using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallbackStars: 322kUpdated Claude CodeCodexVerifiedInstall
verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions alwaysStars: 322kUpdated Claude CodeCodexVerifiedInstall