Learn from Session

Extract actionable learnings and generate project-specific customizations. Adapts Claude Code to project patterns over time.

Use TaskCreate to track these 8 phases:

1. Discover existing customizations
2. Extract learnings from conversation
3. Categorize by artifact type
4. Distill to concrete artifacts
5. Deduplicate & merge with existing
6. Budget check
7. Present & confirm
8. Apply changes

---

Phase 1: Discover Existing

Find ALL project customization files:

# Run these in parallel via Glob
.claude/CLAUDE.md OR CLAUDE.md          # Instructions
.claude/commands/*.md                    # Commands
.claude/skills/*/SKILL.md               # Skills
.claude/settings.json                    # Hooks
.claude/rules/*.md                       # Rules

Record counts for budget tracking later.

---

Phase 2: Extract Learnings

Analyze conversation for these signal types:

Instruction Signals -> CLAUDE.md

Signal	Look For
Corrections	"no", "wrong", "actually...", "instead..."
Direct guidance	"always", "prefer", "use X", "never"
Repeated explains	Same thing clarified 2+ times
Project quirks	Unexpected behaviors, edge cases, workarounds
Commands/workflows	Specific commands, sequences that worked

Command Signals -> .claude/commands/

Signal	Look For	Confidence
Repeated task	Same request 3+ times in session/history	HIGH
"I always" pattern	"I always run X before Y"	HIGH
Multi-step sequence	"first lint, then test, then build"	MEDIUM
Template request	"create a new X like we discussed"	MEDIUM
Workflow mention	"my workflow is X then Y then Z"	HIGH

Skill Signals -> .claude/skills/

Signal	Look For	Confidence
Complex multi-tool	Workflow spanning 3+ files with analysis	HIGH
Agent orchestration	"spawn agents to check X and Y in parallel"	HIGH
Domain expertise	Detailed tech discussion needing reference docs	MEDIUM
Progressive need	"for basic use X, for advanced see Y"	MEDIUM

Hook Signals -> .claude/settings.json

Signal	Look For	Confidence
Blocking rule	"never do X", "always block Y"	HIGH
Validation need	"always run linter after edit"	HIGH
Format requirement	"use prettier on save", "format on edit"	MEDIUM
Approval pattern	"ask before destructive commands"	MEDIUM

If $ARGUMENTS contains topic: Extract ONLY learnings related to that topic. If $ARGUMENTS contains --dry-run: Show what would be created, don't write.

---

Phase 3: Categorize

Sort extractions into buckets:

Instructions: [list of instruction candidates]
Commands: [list of command candidates with names]
Skills: [list of skill candidates with names]
Hooks: [list of hook candidates with events]

Apply confidence threshold: Only include HIGH confidence for commands/skills/hooks. MEDIUM confidence items go to instructions in CLAUDE.md instead.

---

Phase 4: Distill to Artifacts

Instructions -> CLAUDE.md format

Pattern	Example
Use X	Use prepared statements for SQL
Prefer X over Y	Prefer explicit returns over bare err
When X -> do Y	When tests fail -> check CI logs first
X requires Y	Auth endpoints require Bearer token
Run: \...``	Run: kubectl rollout restart deploy/api
Note: X	Note: Get() returns nil, nil when absent

One instruction per line. Max 80 chars.

Commands -> .claude/commands/{name}.md

---
description: { 1-line description }
allowed-tools: { tools needed, e.g., Bash(make *), Read }
argument-hint: { optional args }
---

# {Command Name}

{Brief purpose}

## Steps

1. {step from conversation}
2. {step from conversation}

## Context

{Any relevant @ references or bash context}

Skills -> .claude/skills/{name}/SKILL.md

---
name: { kebab-case-name }
description: { when to use, triggers }
user-invocable: true
context: fork
allowed-tools: { restrictions if any }
---

# {Skill Name}

{Purpose}

## When to Use

- {scenario 1}
- {scenario 2}

## Workflow

1. {step 1}
2. {step 2}

Hooks -> settings.json addition

{
  "EventType": [
    {
      "matcher": "ToolPattern",
      "hooks": [
        {
          "type": "command",
          "command": "command-here",
          "timeout": 30
        }
      ]
    }
  ]
}

Default to PostToolUse (non-blocking) unless explicit blocking requirement.

---

Phase 5: Deduplicate & Merge

For each artifact type:

1. Read existing if file exists
2. Find overlaps by topic/name
3. Merge strategy:

Scenario	Action
New adds specificity	Update existing with concrete details
New adds example	Append example to existing
Both have value	Combine into single comprehensive item
New is subset of existing	Keep existing, skip new
Name collision (cmd/skill)	Merge content into existing file

Prefer updating existing over creating new.

---

Phase 6: Budget Check

Recommended limits:

Artifact	Limit	Why
CLAUDE.md	200 lines	Context efficiency
Commands	10	Discoverability
Skills	5	Complexity management
Hooks	5	Debugging simplicity

If exceeding budget:

1. Propose consolidations
2. Identify stale items for removal
3. Warn but allow user to proceed

---

Phase 7: Present & Confirm

Show ALL proposed changes:

## Proposed Changes

### CLAUDE.md (instructions)

Target: `.claude/CLAUDE.md`
Lines: X/200

**Updates:**

- [old] -> [merged]

**Additions:**

- New instruction

### Commands

- `/pre-commit` - Run checks before commit
  -> `.claude/commands/pre-commit.md`

### Skills

~ `auth-patterns` - Updated with OAuth2 flow
-> `.claude/skills/auth-patterns/SKILL.md`

### Hooks

- PostToolUse[Edit|Write]: Run prettier
  -> `.claude/settings.json`

---

Rollback: `git checkout .claude/`

STOP:

Use AskUserQuestion:

Header	Question	Options
Action	Apply these customizations?	Apply all / Select items / Edit first / No

If "Select items": Show multi-select for each category.

---

Phase 8: Apply

Based on confirmation:

1. CLAUDE.md: Edit existing lines, add new at appropriate sections
2. Commands: Write new .md files, edit existing for merges
3. Skills: Create directory + SKILL.md, add reference files if needed
4. Hooks: Merge into existing settings.json hooks object

Section Placement (CLAUDE.md)

Section	Content
Patterns	Architecture, naming, structure
Code Style	Formatting, idioms, conventions
Workflows	Build, test, deploy commands
Gotchas	Edge cases, quirks, known issues

---

Output

LEARNED
=======
Target: .claude/

Instructions: (CLAUDE.md)
  ~ "Use error wrapping" -> "Wrap errors: fmt.Errorf(\"op: %w\", err)"
  + "Note: API returns nil, nil when absent"

Commands: (commands/)
  + /pre-commit - Run lint, test, build

Skills: (skills/)
  + auth-patterns - OAuth2 and JWT handling

Hooks: (settings.json)
  + PostToolUse[Edit]: prettier --write

Budget: 45/200 instructions | 3/10 commands | 1/5 skills | 1/5 hooks

---

Edge Cases

• Empty session: "No learnings detected. Try after a longer conversation."
• All low confidence: Convert to CLAUDE.md instructions instead
• No .claude/ directory: Create it with user permission
• --dry-run flag: Show output but skip Phase 8

---