reflect

Files (1)

• SKILL.md

SKILL.md (first lines)

# Reflect — Session Learning Loop

## Trigger
`/reflect`

## Purpose
Analyze the current conversation, extract what worked and what didn't, and persist improvements for future sessions. This is how I get better over time.

## Workflow

### Step 1: Session Analysis
Review the full conversation and identify:
- **Tasks completed**: What did the user ask for? Did I deliver?
- **Mistakes made**: Where did I give wrong information, take the wrong approach, or waste time?
- **Corrections received**: What did the user correct me on? (These are the highest-signal learnings)
- **New knowledge gained**: Facts, preferences, patterns discovered this session
- **Approaches that worked well**: What should I do more of?

### Step 2: Classify Learnings
Sort each learning into one of these categories:

| Category | Persists to | Example |
|----------|------------|---------|
| **Behavioural** | CLAUDE.md | "User wants me to stop hedging" |
| **Project fact** | MEMORY.md | "CG converges in 167 iters at N=64" |
| **Bug fix / pattern** | memory topic file | "Windows encoding issue with nlm login" |
| **Skill gap** | New skill or skill update | "User frequently asks for LaTeX tables" |

### Step 3: Update Files
For each learning:

1. **Check for duplicates** — Read the target file first. Don't write what's already there.
2. **Check for contradictions** — If new info contradicts existing memory, UPDATE the old entry, don't add a second conflicting one.
3. **Apply the update**:
   - CLAUDE.md: Add/modify behavioural instruction (keep total under 50 lines)
   - MEMORY.md: Add/modify fact (keep total under 200 lines)
   - Topic files: Create or update as needed
   - Skills: Modify SKILL.md if a skill needs improvement

### Step 4: Report
Show the user exactly what was updated:

```
## Session Reflection

### What I learned
- [learning 1]
- [learning 2]

### Files updated
- CLAUDE.md: [what changed and why]
- MEMORY.md: [what changed and why]
- [other files]: [what changed and why]

### No action taken on
- [things observed but not significant enough to persist]
```

## Rules
- NEVER add speculative or unverified information to memory
- NEVER bloat files — every line must earn its place
- If CLAUDE.md exceeds 50 lines after update, consolidate — but the merged version MUST contain ALL information from originals

### HARD CONSTRAINT: Never degrade
- Only add or improve — never delete working instructions, memories, or user preferences
- Never weaken an existing rule or remove a user correction
- Never replace specific instructions with vague ones
- If unsure whether a change improves or degrades — don't make it, flag it for the user
- Monotonic improvement only — the ratchet only turns one way
- If MEMORY.md exceeds 200 lines, move details to topic files and link from MEMORY.md
- User corrections are the HIGHEST priority learnings — always persist these
- Be honest in the report — don't hide mistakes

From rafeeqinea/claude-academic-toolkit, MIT.

More from rafeeqinea/claude-academic-toolkit