Synthesize accumulated observations into system evolution proposals. Fires when observation count reaches threshold. 5-phase process: triage, methodology updates, pattern detection, proposal generation, and human approval. Never auto-implements — always proposes changes for review. Triggers on: "rethink", "synthesize learnings", "evolve system", "improve process"
Scanned 9/6/2026
Install to Claude Code
npx -y skills add Miosa-osa/canopy --skill rethink --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Rethink?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/miosa-osa-rethink-canopy)More formats (shields.io, HTML) on the badges page.
---
name: rethink
description: >
Synthesize accumulated observations into system evolution proposals. Fires when
observation count reaches threshold. 5-phase process: triage, methodology updates,
pattern detection, proposal generation, and human approval. Never auto-implements —
always proposes changes for review.
Triggers on: "rethink", "synthesize learnings", "evolve system", "improve process"
---
# /rethink
> Synthesize observations into system evolution — always with human approval.
## Purpose
The synthesis side of the learning loop. When enough observations accumulate in a category (tracked by `/remember`), rethink analyzes them for patterns, generates concrete improvement proposals, and presents them for human approval. This is how the system evolves: not through ad-hoc fixes, but through deliberate synthesis of accumulated evidence. Critical rule: rethink NEVER auto-implements changes. Every proposal requires explicit human approval.
## Usage
```bash
# Rethink a specific category
/rethink "process"
# Rethink with confidence override
/rethink "people" --force
# Rethink all categories that are ready
/rethink --all
# Dry run — show what would be proposed
/rethink "technical" --dry-run
# View past rethink proposals
/rethink --history
# Set custom threshold
/rethink "tool-use" --threshold 3
```
## Arguments
| Flag | Type | Default | Description |
|------|------|---------|-------------|
| `<category>` | positional | — | Category to synthesize (from `/remember` categories) |
| `--all` | flag | false | Process all categories that meet threshold |
| `--force` | flag | false | Override confidence threshold |
| `--threshold` | int | 5 | Minimum observations required |
| `--dry-run` | flag | false | Show proposals without creating approval requests |
| `--history` | flag | false | View past rethink proposals and their status |
| `--depth` | enum | `standard` | `quick` (pattern summary), `standard` (proposals), `deep` (proposals + impact analysis) |
## Workflow
### 5-Phase Process
1. **Triage** — Load all observations in the target category. Filter by confidence (>= 0.3). Sort by severity and frequency. Discard noise (low confidence + single occurrence).
2. **Pattern detection** — Cluster related observations. Identify:
- Recurring themes (same mistake, different contexts)
- Escalating severity (problem getting worse)
- Root causes (multiple symptoms, one cause)
- Contradictions (observations that conflict — flag for human resolution)
3. **Methodology update** — For each detected pattern, determine what system artifact should change:
- CLAUDE.md rules
- Agent system prompts
- Workflow definitions
- Skill parameters
- Routing tables
- Quality thresholds
4. **Proposal generation** — For each proposed change:
- Write a specific, actionable proposal
- Include: what changes, why (evidence from observations), expected impact, risk level
- Show the exact diff (before/after) when possible
- Estimate effort: trivial (config change), moderate (workflow change), significant (architecture change)
5. **Human approval** — Submit proposals via `/approve` as `strategy_proposal` type. Block until resolved. On approval, provide implementation instructions (but do not auto-implement).
## Output
```markdown
## Rethink Report — Category: process
**Observations analyzed:** 7
**Patterns detected:** 3
**Proposals generated:** 2
### Pattern 1: Duplicate Content Creation
**Evidence:** 4 observations (confidence: 0.72)
- obs-a1b2: "Always check for duplicates before inserting"
- obs-c3d4: "Created duplicate signal file for Ed's call"
- obs-e5f6: "Knowledge base has 3 versions of pricing decision"
- obs-g7h8: "Ingestion created duplicate when slightly different wording"
**Root cause:** No deduplication check in the ingestion pipeline.
**Proposal 1:** Add dedup check to /seed skill
- **Change:** Add content hash comparison in seed workflow step 5
- **Diff:**
```
Before: 5. Write — Create inbox item
After: 5. Dedup — Hash content, check against existing seeds. If >85% similarity, warn and link instead of creating new.
6. Write — Create inbox item
```
- **Impact:** Prevents ~60% of duplicate content (based on observation frequency)
- **Risk:** Low — additive change, no existing behavior modified
- **Effort:** Moderate — requires hash index
→ Submitted as approval req-x1y2. Awaiting human decision.
### Pattern 2: Missing Cross-References
**Evidence:** 3 observations (confidence: 0.58)
...
### Archived Observations
The following observations were consumed by this rethink and archived:
- obs-a1b2, obs-c3d4, obs-e5f6, obs-g7h8 → archived to ops/observations/process/archived/
```
## Dependencies
- `/remember` — Source of observations
- `/approve` — Proposal approval gate
- Observation store (`ops/observations/`)
- System configuration files (for generating diffs)
- `/health` — For impact analysis at `deep` depth
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!