/retrospective — find one piece of leverage, ship it

One interview = one artifact. Also installs the [[operating-principles]] habits over time. Reads context/priorities.md, context/current-focus.md, connections.md, decisions/log.md, recent daily/*, and existing skills/*.

Phase 1 — Notice (find the candidate)

Ask, conversationally:

1. What did you do 3+ times this week? (frequency)
2. What felt manual, boring, or copy-paste? (drudgery)
3. Where did you think "a smart intern could do this"? (delegation)
4. If your workload doubled tomorrow, what breaks first? (constraint)
5. What knowledge did you keep re-looking-up? (a missing note)

Output: 1-3 candidates, one line of "why this is leverage" each. Ask the user to pick one.

Phase 2 — Decide (scope it)

Walk the five filters from [[operating-principles]]:

1. Constraint — which bottleneck does this solve?
2. Eliminate → Automate → Delegate. Ask "what if we just stop doing this?" If nothing breaks, exit cheerfully — that's a win. Log it to decisions/log.md and stop.
3. Map the steps — trigger, inputs, transformations, decision points, destination. If they can't articulate them, send them to sketch it first.
4. Lowest autonomy that works — suggest < draft < supervise < auto. Default low; push back on "fully autonomous" without proof.
5. Tie to an outcome — more output / less time / fewer errors / learning + a specific metric. If they can't name one, stop.

Write the scoped spec to decisions/log.md as a dated entry.

Phase 3 — Build (ship one artifact)

Ask how to ship, lowest-infrastructure first:

1. A note — a permanent note in notes/ (often the answer for "I keep re-looking this up").
2. A template — a reusable scaffold in templates/.
3. A saved prompt / SOP — a resources/ doc the user runs by hand.
4. A new ritual — a skills/<name>/SKILL.md (deterministic first; one AI step only if needed).
5. A scheduled task — only once the manual version works.

Default to the highest non-automated option that solves it. Build it, confirm it works, and remind the user to run it manually for a week before trusting it.

Output every run

1. One decisions/log.md entry. 2. One shipped artifact. 3. A one-screen recap.

Rules

• Notice always runs first, even if the user arrives with an idea.
• Eliminate-first: if the answer is "stop doing it", that's success, not failure.
• Boring is beautiful — bias to notes/templates/prompts over agents.

When the floor is not met

Notice → Decide → Build asks the model to survey everything, pick the one piece of leverage, and ship it. The survey is the part that degrades: a weaker setup picks the most recent thing rather than the most leveraged, and the ritual becomes a to-do list.

Below the floor: run /audit-notes score instead. It scores the four layers against fixed criteria and names the top three gaps by a stated weighting, so the ranking comes from the rubric rather than from judgment.