| name | coach-meta-review |
| description | Quarterly review where each coach reviews its own CLAUDE.md against actual session data and proposes self-corrections. The coach asks "is my own design still right for this user?" Triggers manually or quarterly. This is Layer 3 of coach evolution. |
Skill: coach-meta-review
Coach-evolve refines content (vocabulary, profile, sessions). Coach-meta-review refines the coach itself — the persona, the protocols, the entire CLAUDE.md.
This is the deepest level of self-improvement: the coach reads what it tells itself to do, compares to what's actually been working, and proposes its own redesign.
When to invoke
- User: "meta-review the pole coach" / "is my coach design still right?"
- Auto-suggested every quarter (~90 days) for coaches with >30 sessions
- Called by
weekly-review skill at quarter boundaries
- After major life changes (injury recovery, new goal, schedule change) — user-triggered
What it reviews
Reads:
- The coach's own
CLAUDE.md (the design under review)
- Last 90 days of
sessions.md
- Last 90 days of phone inbox files (if not yet processed/consolidated)
- Recent snapshot evolution (compare oldest snapshot in last 90 days to newest)
- User's effective vs. ineffective patterns (which sessions led to wins; which to plateaus)
Specific questions the meta-review asks
Persona alignment
- Does my stated voice match my actual messages?
- Are the cues I prescribe the ones the user responds to?
- Have I been adapting to user style, or am I stuck in the original template?
Protocol effectiveness
- Are my "files to read at session start" actually being read and used?
- Does my "files to update at session end" capture what matters?
- Have I missed context-rich files I should be reading?
Threshold tuning
- My CLAUDE.md says "always train left side first" — is the user actually doing this? If not, why?
- My program.md prescribes 30 min of conditioning twice a week — is that realistic given session reality?
- Are my speed budgets (snapshot size, sessions count) right?
Section drift
- What sections of CLAUDE.md haven't been used in 90 days? (Candidates for removal)
- What patterns in the user's data have no corresponding section? (Candidates for addition)
Goal drift
- The user's goals when this coach was spawned — still valid?
- New goals that emerged in conversations but aren't reflected in profile/program?
- Are we training for the right thing?
Output format
## Coach Meta-Review — Pole Coach (2026-11-15, 3 months in, 47 sessions)
### Persona drift (2 findings)
1. **Voice section says "direct, technically precise"** — actual successful messages
are "warmer than the template, with more validation before correction."
Proposed update: soften voice section. (Specific edit shown below.)
2. **"Don't summarize at end of every message"** — you've actually responded well to
short ~3-line summaries when sessions are long. Soften this rule?
### Protocol drift (1 finding)
1. CLAUDE.md says "read last 5 entries of sessions.md" — but recent sessions have
become longer; reading 5 entries puts >30 KB in context. Switch to "read last
3 entries fully + last 10 titles only"?
### Threshold misalignment (1 finding)
1. Program says "always train left side first" but session data shows left-first
training in only 40% of sessions. Either: (a) update program to be more
forgiving, or (b) check whether you're still bought into this rule.
### Section drift (1 finding)
1. The "Cross-coach awareness" section has never produced a cross-coach
suggestion in 47 sessions. Drop it from CLAUDE.md (cleaner) or keep as
aspirational?
### Goal drift (1 finding)
1. Original goal: "shoulder mount clean from the floor."
Current trajectory: progressing well; on track for ~month 5.
Open question: what's the next 6-month goal? Aysha? Handspring? Performance?
### Stable / no change (3)
- Reading protocol files: working as designed.
- Update protocol: working as designed.
- Update behavior at session end: well-calibrated.
Reply with proposed edits to apply (numbers), or "discuss <number>" to talk through one.
Algorithm
- Verify coach has enough data (>30 sessions, >90 days old)
- Read CLAUDE.md, parse into sections
- Read 90-day sessions and inbox history
- For each section of CLAUDE.md, ask: does the data validate this rule?
- For patterns in the data, ask: is there a CLAUDE.md section that addresses this?
- Compose findings; for each, propose a specific edit (with diff-style before/after)
- Show user; wait for input
- Apply approved edits to CLAUDE.md
- Regenerate snapshot via
vault-push-snapshot
- Update
~/.cache/vault-coach-meta-review-history.json
Sample edit format
When proposing a CLAUDE.md edit, show concrete before/after:
## Voice
- Direct, technically precise, encouraging without coddling
+ Warmer than direct — validate effort before correction. Technically precise
+ when needed; conversational otherwise. Match the user's energy.
Don't ask the user to redraft; do the redraft and ask for approval.
Don't
- Don't run before 30 sessions or 90 days (not enough data)
- Don't propose >5 edits per review (overwhelming)
- Don't rewrite the entire CLAUDE.md (incremental edits only)
- Don't be self-deprecating ("I've been failing you...") — just propose changes
- Don't propose changes that conflict with
_shared/principles.md without flagging
- Don't apply edits without explicit user approval
Why this matters
A coach that can't adjust its own design is a static template, no different from examples/ in the public repo. Quarterly meta-review is what keeps your coach genuinely yours — drifting toward what works for you specifically, not what worked at template-design time.
This is also the slowest improvement loop. Per-conversation extraction is fast (every session). Coach-evolve is medium (every 10 sessions). Coach-meta-review is slow (quarterly). Three timescales catch different patterns.
