auto-skill-creator

Automatically creates new skills from complex problem-solving sessions. Triggers at the end of multi-step tasks that required novel approaches, debugging sessions, or workflow creation. Use when: a t

SKILL.md

GitHub

Auto Skill Creator

Turns complex problem-solving into reusable skills — automatically.

When to Trigger

After completing a multi-step task, evaluate:

  1. Complexity — Did it take 3+ tool calls or require debugging?
  2. Novelty — Was the solution non-obvious or undocumented?
  3. Recurrence — Will this pattern likely happen again?

If 2 of 3 are true → create a skill.

Process

1. Extract the Pattern

From the completed task, identify:

  • Problem class — what category of problem was this? (e.g., "proxy crash recovery", "WhatsApp reconnect")
  • Key steps — the minimal sequence that solves it
  • Gotchas — what didn't work or was misleading
  • Prerequisites — what tools/access/knowledge is needed

2. Check for Duplicates

Before creating:

grep -rl "<problem-class-keywords>" skills/*/SKILL.md

If a similar skill exists, update it instead of creating a new one.

3. Create the Skill

Structure:

skills/<skill-name>/
├── SKILL.md          # Frontmatter + instructions
├── scripts/          # Only if deterministic code was part of the solution
└── references/       # Only if reference material is needed

Rules:

  • name: lowercase, hyphenated, verb-led (e.g., debug-litellm-proxy, recover-whatsapp-gateway)
  • description: max 1024 chars, includes what + when to use
  • Body: <500 lines, imperative form, concrete steps
  • Include the gotchas — what looked right but wasn't
  • No README.md, no CHANGELOG.md, no auxiliary files

4. Register in Manifest

After creating the skill directory:

cd /path/to/workspace/skills
python3 -c "
import json, os, glob
manifest = []
for skill_dir in sorted(glob.glob('*/SKILL.md')):
    name = os.path.dirname(skill_dir)
    with open(skill_dir) as f:
        content = f.read()
    # Extract description from frontmatter
    if '---' in content:
        fm = content.split('---')[1]
        for line in fm.split('\n'):
            if line.strip().startswith('description:'):
                desc = line.split(':', 1)[1].strip().strip('\"').strip(\"'\")
                manifest.append({'name': name, 'description': desc[:200]})
                break
with open('_manifest.json', 'w') as f:
    json.dump(manifest, f, indent=2)
"

5. Git Commit

git add skills/<skill-name>/ skills/_manifest.json
git commit -m "auto-skill: <skill-name> — <one-line description>"
git push

6. Log It

python3 tools/eval_tracker.py log proactive_action "auto-created skill: <skill-name>" 2

And add to daily notes:

[HH:MM IL] Auto-created skill <skill-name> from <task description>.

Quality Gate

Before committing, verify:

  • Frontmatter has name and description
  • Description explains both WHAT and WHEN
  • Body is <500 lines
  • Steps are concrete (not "be careful" or "check things")
  • Gotchas section exists if there were false leads
  • No duplicate of existing skill

Anti-Patterns

  • Don't create skills for one-off tasks that won't recur
  • Don't create skills that are just "read this doc"
  • Don't create skills for things the model already knows (basic git, basic Python, etc.)
  • Don't create skills during time-sensitive tasks — queue it and create after

Integration with proactive-pa

The proactive-pa skill's "Next step without asking" pattern should evaluate:

"Did I just solve something hard? Should this be a skill?"

If yes, trigger auto-skill-creator before closing the task.