Weekly Automation Review

Weekly Automation Review

Once a week, look back at recent work, spot the tasks you did by hand more than once, and recommend 1–2 of them to automate into a skill, hook, or tool. The discipline is to automate deliberately on a cadence instead of only when something annoys you in the moment.

Automation is a one-way ratchet unless something also removes it. So each run also does a Retire pass: pick 1–2 existing scaffolds whose capability bet the current model has outgrown and propose deleting them. Adding without ever retiring is how a harness accretes dead weight that fights a better model instead of riding it.

What it reviews

• The prompt log — ~/.claude/automation-review/prompts.jsonl, written by the dev-hooks plugin's prompt-log hook: one JSON line per user prompt (ts, cwd, session_id, len, prompt ≤ 500 chars) across all repos. This is the only cross-repo source, and the strongest signal — what you actually keep asking for is exactly what's worth automating. Pull the last 7 days (jq-only cutoff, so no GNU/BSD date divergence). Read it corruption-tolerantly — a crash mid-append can leave a block of NUL bytes that makes a plain jq -c abort at that line and silently drop the whole rest of the file (so the review sees zero recent prompts and skips its strongest signal). Strip NULs, parse each line on its own, and drop only the unparseable fragment:

  for f in ~/.claude/automation-review/prompts.jsonl ~/.claude/automation-review/prompts.jsonl.1; do
    tr -d '\000' <"$f" 2>/dev/null \
      | jq -Rc 'fromjson? | select(.ts >= (now - 7*86400 | strftime("%Y-%m-%dT%H:%M:%SZ")))'
  done
  

  (prompts.jsonl.1 is the rotation spillover; the loop skips it silently when absent. -R reads each line as a raw string and fromjson? yields nothing for a line that won't parse, so one bad line can never blind the rest.) Cluster prompts by intent — same verbs/nouns ("deploy", "fix CI", "regenerate fixtures"), same cwd, near-duplicate phrasing — and count repeats; anything asked 3+ times in a week, or weekly across reports, is a prime candidate. If neither file exists (the dev-hooks plugin isn't installed, or DEV_HOOKS_PROMPT_LOG=false), the loop prints nothing — skip this source silently and rely on the others.
• Every repo's memory index — glob ~/.claude/projects/*/memory/MEMORY.md and skim each (follow into an individual memory file only when a line looks automation-relevant). feedback/project memories record recurring corrections and workflows ("always run X after Y") — patterns worth turning into a hook or skill. Glob may match nothing; degrade silently.
• Recent git activity — commits/branches in the working repo over the last 7 days (git log --since='7 days ago' --stat); what kinds of changes repeated?
• The off-topic backlog — plans/off-topic-improvements.md if present; recurring themes there are automation candidates.
• Repetition signals — the same multi-step manual workflow done more than once, the same class of fix, the same checklist run by hand.
• The previous report — read the last file in plans/automation-reviews/ so recommendations build on prior ones and you can note what actually got automated (self-improving loop).

Procedure

1. Read the previous report (if any) and note which past suggestions shipped.
2. Gather the week's activity — start with the prompt log (when present), then the memory indexes, then git/backlog.
3. Identify repeated workflows — favor ones that are high-volume, well-defined, low blast-radius, and measurable.
4. Pick the top 1–2 and, for each, specify: what it is, whether it's best as a skill / hook / tool, a rough effort estimate, and the first concrete step.
5. Retire pass (below) — pick 1–2 scaffolds to propose removing.
6. Write the dated report (below) and post a one-paragraph summary to the user.

Retire pass — prune scaffolding the model outgrew

The counterweight to the add pass. A hook or skill that compensates for a model weakness declares its bet in its header (# bet: <weakness> / # sunset: <observable that means delete me> for hooks; a bet:/sunset: note for capability-compensating skills). A bet of none (L1–L6) means it encodes preference / private fact / verification / safety — those don't depreciate, so skip them. Focus only on the ones making a real capability bet.

Each run, take 1–2 such scaffolds (rotate — favour the oldest, or ones the fire log shows firing rarely) and check the bet against the current model:

• Evidence. If ~/.claude/automation-review/hook-fires.jsonl exists (opt in with DEV_HOOKS_FIRE_LOG=1), see how often the hook actually fires. A hook that never fires is either dead or its trigger is stale; a hook that fires constantly and is always ignored is noise.
• The sunset test. Read the # sunset: line — it names the observable that means "delete me." Judge whether the current model now does that thing unprompted (sample a few recent transcripts, or run a quick with/without probe on a representative case).
• If the bet no longer holds, propose removal (or downgrade a blocking gate to advisory) with the same PR discipline as an add: one change, tests updated, the reasoning recorded.

Run this pass explicitly on every major model release, not only weekly — a capability jump is exactly when standing bets go stale in bulk.

Output

Write to plans/automation-reviews/YYYY-MM-DD.md:

# Automation Review — YYYY-MM-DD

## Reviewed
<repos / window / sources scanned>

## Repetitive workflows spotted
- <workflow> — seen <N>× — <why it recurs>
- <request cluster> — asked <N>× across <M> repos (prompt log) — <why it recurs>

## Recommended this week (1–2)
1. <name> — skill | hook | tool — effort: <S/M/L>
   - First step: <…>

## Retire this week (0–2)
- <hook/skill> — bet: <its bet> — verdict: <still holds / model outgrew it> — <keep / delete / downgrade>

## Status of prior suggestions
- <past idea> — shipped / dropped / still open

Then give the user a short summary + the top recommendation.

Scheduling

Registered as a recurring Monday-morning remote agent via the [[schedule]] skill (CronCreate). Note: remote routines run against a single repo, and a remote run can't see the local prompt log or the per-repo memory indexes (they live on your machine under ~/.claude/) — so the cross-repo prompt-log/memory view only materializes on a local run. For a genuine sweep across all local projects, run it via a local recurring mechanism (/loop or a local cron) instead. Each run appends to the report history, so the review compounds over time.