Worked example

Spec 38 is one feature carried through the whole pipeline in Gomi Bunrui, a bilingual (English and Japanese) waste-sorting app for residents in Japan. The files are copied without rewriting, and the example’s README records the one path substitution made for publication.

Note

The artefacts were produced with an earlier snapshot of the skills, dated 2026-04-25. The skills have changed since then, so read the skill pages for current behaviour. Use this example to see the shape of each document.

The feature

Every AI classification answer must cite the municipal rule it draws from and carry a last-verified date. The cadence for re-verifying those rules is published as an accuracy-update SLA. When a municipality has not been re-verified inside the SLA, the answer shows a “not recently verified” state instead of hiding the gap.

The work spans a React component, a Kotlin Lambda service, a new DynamoDB audit table, an IAM-protected admin route and bilingual copy.

The artefacts in pipeline order

Step 1: brainstorm.md

brainstorm.md was written on 2026-05-13 by superpowers’ brainstorming skill and copied into the work item by spec-generator. Today spec-discovery does this step and writes discovery.md.

Step 2: requirements.md

requirements.md is spec-generator Phase 1, written on 2026-05-13 and amended on 2026-05-23. Read it for the EARS acceptance criteria that every later document traces back to.

Step 3: design.md

design.md is spec-generator Phase 2, written on 2026-05-13 and amended on 2026-05-23. Collaborator roles added the Data Architecture, Security Architecture and Copy Intent sections.

Step 4: design-review-summary.md

design-review-summary.md is the spec-design-review gate, run on 2026-05-13. The verdict was PASS with four MINOR findings, MINOR-001 to MINOR-004, and no CRITICAL or MAJOR. Start here if you want to see what a gate report looks like.

Step 5: tasks.md

tasks.md is spec-generator Phase 3. Its Status Summary table was updated during implementation.

Step 6: no task review

The spec-task-review gate was not run on this feature, so there is no task-review-summary.md. One visible result: brainstorm.md is still present, because in that snapshot the gate deleted it on a pass. Today the gate deletes discovery.md when it returns APPROVED or APPROVED WITH ISSUES.

Step 7: test-plan.md

test-plan.md is spec-qa-review in plan mode, written on 2026-05-13.

Step 8: implementation

Implementation ran through superpowers and left no file in this directory.

Step 9: implementation-audit.md

implementation-audit.md is the spec-implementation-audit gate, run on 2026-05-23. The verdict was PASS with four MINOR findings, MINOR-005 to MINOR-008, all drift between spec and code in naming and wording. Every commit hash claimed in tasks.md was verified against the branch.

Step 10: qa-review.md

qa-review.md is spec-qa-review in review mode, run on 2026-05-23. The verdict was PASS with one new MINOR finding, MINOR-009. The 85 percent unit-coverage gate was not met on the spec’s own classes, and the reviewer classified that as MINOR because the end-to-end suite was green.

Step 11: spec-maintenance-report.md

spec-maintenance-report.md is spec-maintenance, run on 2026-05-23. Six MINOR findings were resolved by amending requirements.md, design.md and tasks.md, and three earlier specs were annotated.

Step 12: the pull request

The pull request left no file in this directory.

Where the artefacts differ from the skills

The README reports these differences and does not reconcile them:

1. There is no task-review-summary.md, because step 6 was skipped.
2. qa-review.md is a separate file. The review mode is meant to update test-plan.md in place, and here test-plan.md still reads “Pending” throughout.
3. requirements.md, design.md and tasks.md are the versions after maintenance. Step 11 amended them in place and marked each amendment inline. The originals from steps 2 and 3 are not kept.
4. The checkboxes in tasks.md disagree with its Status Summary. Tasks 1 to 18 are unticked while the summary marks them done with commit hashes. Today’s implementation audit would fail those tasks, because only a ticked box completes a task.
5. Role bylines, AIW-nnn issue links and the hand-off blocks come from the orchestration layer of the private project, not from the skills. The AIW-nnn links do not resolve here.

No gate on this feature returned BLOCKED, so the example does not show a blocking verdict.