| name | content-engine-review |
| description | Monthly retrospective on the content engine. Reads last 30 days of calendar-log, memory, grade history, and pickup rates. Surfaces what worked, what didn't, proposed tuning. Saves report to reviews/YYYY-MM-review.md. Foundation B + C applied. |
| when_to_use | ["/content-engine-review","monthly review","engine review","content retrospective","content retro","what worked last month","review my content engine"] |
/content-engine-review — v0.1.0
Monthly retrospective on the cowork-social engine. Runs ~10 minutes. Surfaces what worked, what didn't, and proposes specific tuning before next month starts.
Live skill (not headless). Walks the user through findings + asks decisions on staged skill changes. ⚡ NEXT MOVE prints in chat, varies by verdict (3 states).
Inputs (lazy-load — Foundation A)
Read on every run:
| File | Why |
|---|
projects/social-media-content/calendar-log.md (last 30 days slice) | Posts shipped, platform breakdown, scheduled vs manual, status |
projects/social-media-content/memory.md (last 30 days slice) | Self-improvement notes, recurring complaints, headless-run logs |
projects/social-media-content/skill-improvements.md | Staged changes waiting on user review |
projects/social-media-content/ideas/ (last 4 weekly idea files) | Generated vs picked vs shipped funnel |
projects/social-media-content/brand-brief.md | Current brand state — is the wedge stale? |
_aibos/state-social.md | blotato_status, selected_platforms, target cadence |
_shared/foundations.md | Foundation B + C |
If calendar-log.md has fewer than 5 rows in the last 30 days → surface:
"Less than 5 posts shipped in the last 30 days. The retrospective will be light. Continue, or block 2 hours for /weekly-content-session first?"
If user says "block 2 hours" → halt. If "continue" → proceed with whatever data exists.
The 6-step flow
Step 1 — Pull the data
Aggregate from each input file:
From calendar-log.md (last 30 days):
- Total posts shipped (count)
- Per-platform breakdown (count + % of total)
- Status breakdown: scheduled / published / manual / drafted / failed / skipped
- Day-of-week distribution
- If Blotato connected → success rate = published / (scheduled + published + failed)
From memory.md (last 30 days):
- All self-improvement notes per skill
- All headless-run logs from
/generate-weekly-ideas
- All failure notes
From skill-improvements.md:
- All rows where
reviewed: no — these are staged changes the user hasn't yet decided on
From ideas/<last 4 weekly files>:
- Ideas generated: total = 40 (4 weeks × 10)
- Ideas picked: cross-reference
calendar-log.md against idea hooks (substring overlap >= 50%)
- Ideas drafted but not shipped: in
outputs/ but not in calendar-log.md
- Ideas shipped: in
calendar-log.md with status: scheduled or published or manual
- Compute the funnel: generated → picked → drafted → shipped
Step 2 — Compute patterns
Surface the patterns the user might miss:
Platform performance:
- Top platform this month (most posts)
- Lowest platform (least posts) — flag if it's a
selected_platforms entry
- If grade history is in
calendar-log.md (Step 3 of /weekly-content-session logs grades), compute avg grade per platform
Hook pattern wins:
- Which of the 11 hook patterns appeared most in
outputs/social-media-content/ drafts this month
- Correlate hook pattern → grade score (if grades logged) — name the top 2 patterns by avg score
Cadence adherence:
- Compare actual posts/week vs target (from
_aibos/state-social.md or scheduling-defaults.md)
- Compute adherence %
Self-improvement recurrence:
- For each skill, count how many self-improvement notes appeared in
memory.md
- Flag any pattern recurring 3+ times that's NOT already in
skill-improvements.md — auto-stage per Foundation B
Idea pickup funnel:
- Generated → Picked %: of 40 ideas, how many became picks?
- Picked → Drafted %: of picks, how many became drafts?
- Drafted → Shipped %: of drafts, how many shipped?
- The narrowest stage = the bottleneck
Brand-brief staleness:
- Check
brand-brief.md last_updated field
- If older than 90 days OR no posts referenced
contrarian_belief in their hook pattern → flag
Step 3 — Plan-then-approve the report write
Before writing the report file, show the user a summary table + ask approval:
Last 30 days at a glance:
─────────────────────────
Posts shipped: <N> across <M> platforms
Top platform: <X> (<N> posts, avg grade <X>)
Worst platform: <Y> (<N> posts, avg grade <X>) — flag for tuning
Top hook pattern: <name> (<count> uses, avg grade <X>)
Cadence: <actual>/week vs <target>/week target → <adherence %>
Idea funnel: 40 generated → <N> picked (<X%>) → <N> drafted → <N> shipped
Staged tuning: <N> skill-improvements rows waiting on you
Approve writing the full report to reviews/<YYYY-MM>-review.md?
Reply 'go' to write, or 'skip' to keep this conversation-only.
Plan-then-approve is non-negotiable for the file write. If user says skip → don't write the file, but still apply Foundation B + C in chat.
Step 4 — Write the report
If approved, resolve path: projects/social-media-content/reviews/<YYYY-MM>-review.md.
Create reviews/ directory if missing.
Write:
---
type: cowork-social-monthly-review
plugin: cowork-social
plugin_version: 0.1.0
review_month: <YYYY-MM>
generated_date: <ISO>
posts_shipped: <N>
adherence_pct: <X>
source_skill: /content-engine-review
---
# Content Engine Review — <Month Year>
## Summary
- **Posts published:** <N>
- **Platform breakdown:** linkedin <N>, twitter <N>, instagram <N>, tiktok <N>, threads <N>, facebook <N>
- **Avg posts/week:** <X>
- **Target posts/week:** <X>
- **Adherence:** <X>%
- **Blotato success rate:** % (if connected) | not connected — manual posting can't be measured
# — uses, avg grade /100
— posts, avg grade /100,
— , grade — <1-line why it worked>
— , grade — <1-line why it worked>
— , grade — <1-line why it worked>
posts, avg grade .
entries in waiting review (see Staged section below).
missed scheduled slots.
. % of never became .
Last updated .
| skill | pattern | firstchange |
|---|---|---|---|---|
| / | | | | |
—
—
—
---
Generated by on .
Step 5 — Walk staged improvements with user
For EACH row in skill-improvements.md where reviewed: no, ask the user:
"Apply this change to /<skill>? Pattern: '' recurred times. Suggested change: '<suggested_change>'. Reply Y / N / defer."
Branch:
- Y → mark row
reviewed: yes — approved <ISO>. Open the target SKILL.md for editing (or surface the file path for the user to edit). Append the approval to memory.md.
- N → mark row
reviewed: yes — declined <ISO> | reason: <user's one-liner>.
- defer → leave row as
reviewed: no. It'll show up in next month's review.
If 5+ rows are pending, batch: ask the user "5+ staged. Want to walk all of them now (~5 min) or defer this and only walk the top 3 by recurrence count?"
Step 6 — Brand-brief refresh check
If brand-brief staleness flagged in Step 2, ask:
"Brand-brief was last updated — days ago. <If 3+ self-improvement notes mentioned voice mismatch:> Multiple voice-mismatch notes this month suggest the brief drifted. Run /brand-brief now to refresh? Y / N / defer."
If Y → hand off to /brand-brief (the user's next skill invocation).
If N or defer → log in memory.md and move on.
Foundation B — Self-improvement close
See _shared/foundations.md → Foundation B. After delivering the report + the ⚡ NEXT MOVE block, ask the user:
"What would've made this review 10% better?"
Append the answer to projects/social-media-content/memory.md:
<YYYY-MM-DD> | /content-engine-review | <answer verbatim>
Recurrence patterns to watch for this skill specifically:
- "review took too long" → consider trimming Step 5 walkthrough
- "didn't find the bottleneck" → consider deeper funnel analysis in Step 2
- "recommendations were generic" → consider more business-brain.md context pull
- "missed grade trends" → consider better grade-history aggregation
If any pattern hits 3+ → flag in skill-improvements.md per Foundation B rules.
Foundation C — ⚡ NEXT MOVE block (verdict-aware, 3 variants)
See _shared/foundations.md → Foundation C. Pick the variant based on the month's verdict:
Verdict 1 — Shipping consistently (>= 10 posts/month) AND grade trend flat-or-up
⚡ NEXT MOVE: Schedule a 30-minute /brand-brief refresh this week.
Why: You're shipping consistently — time to refresh the angle so the next 30 days don't recycle the same hooks.
Verdict 2 — Shipping is the bottleneck (< 5 posts/month)
⚡ NEXT MOVE: Block 2 hours Monday for /weekly-content-session.
Why: Last month's shipping rate is below the engine's break-even — habit needs a calendar slot.
Verdict 3 — Grade trend dropping (avg grade lower than prior month by >= 5 points)
⚡ NEXT MOVE: Re-run /brand-brief on the recent_proof_story field this week.
Why: Grade trend is dropping — voice may have drifted from your real proof, hooks are getting generic.
Picking rule (when multiple verdicts apply)
Priority order:
- Verdict 2 wins if shipping < 5/month (no point fixing voice if nothing's going out).
- Verdict 3 wins if grades dropped >= 5 points (voice drift kills the engine even at high volume).
- Verdict 1 is the default when shipping + grades are both healthy.
Validation pattern (same as canonical): ⚡ NEXT MOVE: .+ .+ .+\n Why: .+
If the block doesn't match → regenerate before printing.
Hard rules
- Plan-then-approve before writing the report file. Step 3 must complete before Step 4 fires.
- Don't auto-apply skill changes. Step 5 walkthrough — always confirm each row.
- Reviews directory:
projects/social-media-content/reviews/ — create if missing, never overwrite past months.
- 3 verdict-aware NEXT MOVE variants — never a generic catch-all.
- Apply 3 foundations every run — Foundation A (lazy-load) is automatic; B + C are explicit close.
- Reference
_shared/foundations.md rather than duplicating foundation text.
- 3rd-4th grade reading level in user-facing prompts.
- Surface patterns the user might miss — that's the whole point of the retrospective. Don't flatter, don't hedge.
Voice
The retrospective's voice: direct, opinionated, willing to tell the user a platform isn't working. Not flattery, not hype.
When in doubt: imagine a senior strategist sitting across the table on the first of the month. They've read the data. They have an opinion. They give it in plain words.
