| name | pmf-review |
| description | Roll up the week's daily PMF logs into a Weekly Review entry (Focus / Learned / Decisions / Next), surface recommended hypothesis signal/status updates for you to confirm, and refresh the index. Use at the end of a week, or when the user says 'pmf review', 'do my weekly review', 'roll up the week', 'weekly synthesis', or types '/pmf-review'. |
| argument-hint | [optional: week-start YYYY-MM-DD, and/or path to the pmf/ root] |
| version | 0.1.0 |
| user_invocable | true |
/pmf-review — weekly rollup
You are synthesizing a week of daily logs into one honest weekly review. Roll up, don't inflate:
three n=1 notes on the same point may become n=3, but only say so explicitly and only if justified —
never launder repetition into a trend.
Before you do anything
- Read
${CLAUDE_PLUGIN_ROOT}/references/honesty-rules.md (NON-NEGOTIABLE), format-spec.md, and evidence-weight.md.
- Resolve the tracker root
$PMF and the week window: default to the last 7 days ending today
(date +%Y-%m-%d); if $ARGUMENTS gives a week-start date, use [start, start+6d].
How to run it
There are two equivalent paths. Prefer the workflow if your environment runs workflow scripts; otherwise
do it inline. Both produce the same review.
Path A — workflow (typed, parallel)
The canonical implementation is ${CLAUDE_PLUGIN_ROOT}/workflows/pmf-review.js. It runs four phases —
Collect (resolve root + window, list the week's logs, read the prior review's ## Next), Read
(one sub-agent per daily log, in parallel, extracting did/learned/convos/movement/decisions with evidence
weights preserved — no cross-day synthesis), Synthesize (one agent drafts the four review fields and
justifies any evidence-weight rollup), and Write (write the review file, record recommended
signal/status changes for you to confirm via /pmf-wrap, regenerate the index). It never edits
hypothesis files itself. Invoke it as a workflow if available.
Path B — inline (always works)
- Collect. List
$PMF/log/*.md whose date is in the window. Read the newest $PMF/reviews/*.md's ## Next.
- Read. Read each in-window log. For many logs, fan out one read-only sub-agent per log (Task/Agent
tool) to extract
{what_did, what_learned, customer_convos, hypothesis_movement, decisions} verbatim
with evidence weights — instruct each sub-agent to follow honesty-rules.md and NOT synthesize across days.
- Synthesize. From the extracts + the prior
## Next, draft the four fields:
## Focus — what the week was actually about.## Learned — grounded findings; roll up evidence weight explicitly (show the n=), no invented trends.## Decisions — decisions actually made.## Next — concrete next steps (carry forward unfinished items from the prior ## Next).
- Write. Create
$PMF/reviews/<week-start>.md from ${CLAUDE_PLUGIN_ROOT}/templates/weekly-review.md;
set period, logs_rolled_up (the filenames), and hypotheses_touched (union across the week).
Record any hypothesis signal or status changes the week suggests as recommendations in the
review's ## Next (e.g. "Recommend H4 signal — → Mixed; confirm via /pmf-wrap"). Per
honesty-rules.md rule 4, neither signal nor status changes to hypothesis files happen here without
explicit confirmation — apply them in /pmf-wrap, not in the review. (If you want to apply a change
now, switch to /pmf-wrap, which gates it with AskUserQuestion.)
After writing
Regenerate the index: python3 ${CLAUDE_PLUGIN_ROOT}/scripts/pmf_index.py --out "$PMF" --check. Report the
review path, which hypotheses changed, and that the index refreshed. If the week had zero logs, say so and
write nothing.
