Weekly retro skill — surface candidates for packaging into skills

Once a week, an agent reviews all GitHub Issues closed in the current ISO week (its memory store — see #127) and flags repeated sequences as candidates to be packaged into reusable skills. You get a ranked list: "you did X manually 4 times this week — promote to skill?"

Compounds personal automation. Every week the operator's AI ops get slightly more efficient instead of stagnating at "I keep typing the same prompt."

1. Prerequisite: every task creates a GitHub Issue (see #127 Project-as-Memory).
2. Tag each Issue with the ISO week (e.g. week-19-2026).
3. Sunday cron or manual trigger runs the weekly-retro skill: "review all week-19-2026 issues. Cluster by action pattern. For each cluster ≥3 items, propose a skill name + one-paragraph description."
4. Skim the output. Promote 1–2 top candidates into actual skill files in skills/.
5. The next week's retro should show those former candidates as "now automated."

• Don't promote everything. Three repetitions is the floor; below that you're encoding noise.
• The skill suggestions read smarter than they are — review with skepticism. A skill that needs five prompt-tweaks per run is worse than no skill.