Continuous Learning

Continuous Learning

Scan archived session logs to surface recurring manual workflows, then offer to convert high-frequency patterns into new SKILL.md stubs so repeated work becomes permanent automation.

Usage

/continuous-learning              # Scan all archived sessions for patterns
/continuous-learning --min 2      # Lower threshold (default: 3 repetitions)
/continuous-learning --generate   # Auto-generate stubs without prompting

Behavior

Phase 1: Discover sessions

# Find all archived session files (not the active .current-session pointer)
find .claude/sessions -name "*.md" -not -name ".current-session" 2>/dev/null | sort

If fewer than 3 session files exist, report "Not enough session history yet (need 3+)" and exit.

Phase 2: Extract commands and workflows

Scan each session file for lines that match invocation patterns:

# Commands invoked (slash commands, skill names)
grep -h -E "^/[a-z]|^\*\*Command\*\*:|Ran:|Invoked:" .claude/sessions/**/*.md 2>/dev/null \
  | sed 's|.*: ||' | sort | uniq -c | sort -rn | head -40

Also extract multi-step sequences: look for clusters of 3+ consecutive commands that appear together in at least --min sessions.

Phase 3: Build pattern registry

Load or create .claude/cache/continuous-learning/patterns.json:

{
  "last_scan": "2026-06-04T00:00:00Z",
  "patterns": [
    {
      "commands": ["/security-scan", "/review", "/test"],
      "frequency": 5,
      "sessions": ["session-a.md", "session-b.md"],
      "candidate_name": "quality-gate"
    }
  ]
}

Merge new findings with prior registry. Patterns that drop below --min threshold are removed.

Phase 4: Present findings

Display a ranked table of patterns that meet the threshold:

Repeated patterns found across your sessions:

  Rank  Frequency  Pattern
  ────  ─────────  ───────────────────────────────────────────
  1     7x         /security-scan → /review → /test
  2     5x         /find-todos → /fix-todos → /commit
  3     4x         /db-diagram → /migration-generate → /seed-data
  4     3x         /brainstorm → /write-plan → /implement

Convert any of these to a skill? (enter rank, or 'none'):

Phase 5: Generate skill stub (if requested)

For the chosen pattern, emit a SKILL.md stub using the skillify template:

---
name: 
description: 
disable-model-invocation: false
risk: safe
---

# 

## Usage

\`/\`

## Behavior

1. Run `/step-one`
2. Run `/step-two`
3. Run `/step-three`

## Token Optimization

**Expected range**: 200–600 tokens (delegates to constituent skills)
**Early exit**: Each constituent skill handles its own early exit.
**Patterns used**: Delegation to existing skills

Save to skills//SKILL.md and confirm location to user. Do not commit.

Examples

No patterns yet:

/continuous-learning
→ Scanned 2 session files — need at least 3 to detect patterns.
  Run more sessions, then try again.

Patterns found:

/continuous-learning
→ Scanned 8 sessions. Found 3 repeated patterns (threshold: 3x).
  [table shown]
  Convert #1 to a skill? → User enters "1"
  → Stub written: skills/quality-gate/SKILL.md

Token Optimization

Expected range: 400–900 tokens (initial scan), 150–300 tokens (cache hit)

Caching: Stores pattern registry in .claude/cache/continuous-learning/patterns.json. On subsequent runs, only sessions newer than last_scan are re-processed; old results are merged from cache.

Early exit: Exits immediately if fewer than 3 archived session files exist.

Patterns used: Grep-before-Read (scan with grep, never read full session files into context), Bash for system queries, progressive disclosure (summary table before detail), early exit.

Edge Cases

• No sessions directory: Reports setup instructions and exits.
• Sessions with no command lines: Skips those files silently.
• Candidate name collision with existing skill: Appends -2 suffix and warns user.
• User declines all patterns: Exits cleanly with "Nothing converted."

Safety

• Never deletes or modifies existing skills or session files.
• Skill stubs are written to skills/ only — not installed to ~/.claude/skills/ automatically.
• Does not commit anything; user must run /commit explicitly.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: manastalukdar
• Source: manastalukdar/ai-devstudio
• License: MIT

Install and usage instructions live in the source repository linked above.