session-to-skill

<!-- Adapted from cc-haha/src/skills/bundled/skillify.ts (MIT). Logic: session analysis → user interview → SKILL.md generation. Changes: adapted to yamtam SKILL.md format, added Anti-Fake-Pass gates, removed ANT-only constraints. -->

When to Use

• Use when: user just completed a multi-step workflow they want repeatable
• Use when: user explicitly says "make this a skill" or "I want to automate this"
• Use when: the task has ≥3 distinct steps that follow a clear order
• Do NOT use: for one-off tasks, bug fixes, or exploratory investigations
• Do NOT use: if an equivalent skill already exists — check first

Pre-check

Before starting, run:

ls core/skills/ | grep -i "<proposed-name>"

If a similar skill exists, suggest updating it instead.

4-Step Process

Step 1 — Analyze the session

Read the conversation and identify:

• What repeatable workflow was just performed?
• What were the distinct, ordered steps?
• What inputs did it require (files, URLs, flags)?
• What was the success criteria?

Write a 1-sentence summary before proceeding.

Step 2 — Interview the user (use AskUserQuestion)

Round 1 — Name and scope:

• Proposed skill name (kebab-case)
• One-line description: "Use when..."
• What should it NOT be used for?

Round 2 — Steps and inputs:

• Confirm the step sequence
• What arguments/inputs are required vs optional?
• Any environment requirements (env vars, tools)?

Round 3 — Trigger phrases:

• What would the user type to invoke this? (3–5 trigger phrases)
• What signal words make it obvious this skill applies?

Step 3 — Draft the SKILL.md

Follow the yamtam SKILL.md format:

---
name: <skill-name>
description: >
  <one-paragraph description with trigger phrases and Do NOT cases>
origin: session-extracted
version: 1.0.0
---

## When to Use
## Pre-conditions
## Steps
## Anti-Fake-Pass

Keep it under 200 lines. One clear thing per step. No padding.

Step 4 — Show and confirm

Show the complete draft to the user. Ask: "Does this capture it correctly? Any adjustments before I save?"

Only write the file after explicit approval.

Write to: core/skills/<name>/SKILL.md

Anti-Fake-Pass

❌ Creating a skill for a one-off task with no repeatable value
❌ Writing the skill without confirming the step order with the user
❌ Saving without showing the draft first
❌ Skipping the duplicate-check
✅ User explicitly confirmed the draft before saving
✅ Skill has ≥3 distinct steps
✅ Trigger phrases are specific enough to avoid false positives