tune

tune — turn friction into fixes

The issue-to-pr pipeline no longer has an always-on "improve this skill" step. Instead each run appends one line to .claude/issue-to-pr/friction.log (repo-local, gitignored) whenever a step genuinely fought back — a stop that should not have happened, an unclear instruction, a missed detection, a script that surprised the model. This skill batches that evidence into real changes.

Steps

1. Read the log. .claude/issue-to-pr/friction.log (from the repo root). Each line is a dated, one-sentence friction note with the step it hit. If the file is absent or empty, say so and stop — nothing to tune.
2. Cluster. Group the lines by root cause, not by surface symptom. Three notes about the same confusing exit code are one fix, not three.
3. Propose. For each cluster, propose the smallest concrete change to the SKILL, a script, or a reference — quoting the friction lines as evidence. A change with no evidence line does not belong here. Prefer edits that make a stop mechanical (an exit code / a test) over prose that asks the model to remember.
4. Confirm, then apply. Show the batched proposal and ask for a go-ahead. On approval, make the edits (bump the plugin version + changelog per repo rules), run the gates (tests/run-tests.sh), and — if the changed area is behavioral — note that a fresh dogfood run should confirm it. With --dry-run, stop after the proposal.
5. Prune. After applying, remove the addressed lines from the friction log (keep any not yet acted on) so the log reflects only open friction.

Never apply blind: every edit traces to a friction line, and the human approves the batch. This keeps the skill improving from real use, not speculation.