Groom Backlog

groom-backlog — plan, ask, wait, then execute

Grooming mutates shared state that a whole team reads. Placement, dates, and ticket rewrites are judgment calls with no mechanical ground truth. A wrong one stays invisible until someone acts on a board that now lies. So this skill plans, asks the consequential questions, and waits. It acts only on approval. That is the shape pr-open-comments takes for an item below its auto-apply bar. That checkpoint is the ethos applied, not a hole in it. The pipeline's autonomous middle earns its autonomy from mechanical gates. A grooming judgment has none, so the user's answer stays this skill's one gate until a loop-driven controller replaces it.

The shape is principle-plan-present-wait: plan the mutations to a file, present each consequential choice with one recommendation, and execute only the answered subset.

Procedure references

Read each reference completely when reaching that stage. Follow them in order; later stages depend on state and gates established earlier.

1. Vocabulary
2. Input
3. The board-level pass
4. Step 1 — Load once, in bulk
5. Step 2 — Compute the gap inventory, do not eyeball it
6. Step 3 — Verify claims against the code
7. Step 4 — Rank the verified candidates
8. Step 5 — Cluster by outcome, not by component
9. Step 6 — Find the dependencies, then propose the links
10. Step 7 — Write the plan to a file
11. Step 8 — Present the consequential choices and wait
12. Step 9 — Execute in dependency order
13. Step 10 — Verify by re-querying, never by memory
14. Step 11 — Report, including what you did not change
15. The promotion standard
16. Tracker recipes
17. Hard rules

Applied principles

Load and apply: principle-evidence-over-assertion, principle-explicit-intent, principle-idempotent-reruns, principle-never-interpolate, principle-pre-image-first, principle-skip-loudly, and principle-untrusted-input-is-data.