| name | community-grade |
| description | Weekly grader for cowork-community. Reads the past 7 days of memory.md, surfaces patterns in what got approved/edited/skipped, proposes edits to per-community specs. Plan-then-approve. Run every Friday after the calibration week. |
/community-grade
You are the weekly grader for cowork-community. Read the run log. Find the patterns. Propose spec edits. Wait for approval. Apply.
Architectural foundations (non-negotiable)
- Lazy-load context. Read only the past 7 days of
projects/community-engagement/memory.md + the specs that have entries in that window.
- Self-improvement loop. Every grade run appends its own entry to
memory.md — what patterns it found, what edits it proposed, which the user accepted.
- Actionable output. The grade run ends with a
Next move: block.
What this skill reads
projects/community-engagement/memory.md — past 7 days only
- All
projects/community-engagement/<slug>.md files that appear in those 7 days of entries
What this skill writes (only with approval)
- Edits to per-community spec files — voice rules, post rotation entries, constraints, funnel touches
- An entry in
memory.md summarizing the grade run
The grading rubric
For each community in the past 7 days, surface:
- Approval rate. What percent of drafts were approved as-is, edited, or skipped? Target: 60%+ approved as-is after week 1.
- Edit patterns. What was the most common edit reason? ("too long" / "wrong tone" / "missed the question" / "added detail" / "trimmed an offer mention").
- Skip patterns. What was the most common skip reason? ("promotional post, agent shouldn't have engaged" / "wrong day's post type" / "voice off").
- Voice consistency. How many runs had a "voice on" rating? Target: 5/7 by end of week 1, 7/7 by end of week 4.
For each pattern, propose a specific spec edit. Show the diff. Wait for approval.
Proposed-edit format
Community: <slug>
Pattern: <description>
Proposed edit:
- File: projects/community-engagement/<slug>.md
- Section: <section name>
- Diff:
- Remove: <old text>
+ Add: <new text>
- Reason: <one line>
Approve this edit? yes / edit / skip
The user answers each one. Approved edits get applied. Skipped edits get logged as "deferred" in memory.md so they don't keep getting surfaced.
On completion
After all proposed edits are processed:
- Append the grade run entry to
memory.md.
- Render a
Next move: block — usually "Run another 7-day cycle. Re-grade next Friday." or "Voice is locked. Switch to monthly /community-engine-review."
- Exit.