/session-to-skill — Session Distiller + Skill Factory

You convert the current chat session into a cross-platform agent skill package following the Agent Skills Open Standard (SKILL.md). The process has two stages: summarize, then build.

Trigger

User invokes /session-to-skill optionally followed by a skill name or focus area:

/session-to-skill
/session-to-skill deploy-checklist
/session-to-skill Focus on the CDK patterns we discussed

Stage 1: Session Distillation

Scan the entire conversation history. Extract only action-oriented prompts — messages where the user requested code changes, research, creation, debugging, architecture decisions, or tool usage. Skip informational questions, greetings, and meta-discussion about capabilities.

For each actionable prompt, produce a summary entry:

### Prompt N: <one-line description>
- **Action**: What the user asked for
- **Domain**: The technical area (e.g., CDK, IAM, skill creation, API design)
- **Outcome**: What was produced or decided
- **Pattern**: The reusable workflow or knowledge (if any)

After listing all entries, synthesize two things:

1. Workflow patterns — Recurring sequences of actions (e.g., "research repo → analyze structure → build based on findings"). These become the skill's procedural instructions.
2. Domain knowledge — Facts, conventions, formats, and decisions accumulated during the session (e.g., "Kiro skills use YAML frontmatter with name and description fields"). These become the skill's reference material.

Present the distillation to the user and ask:

• Does this capture the right scope?
• Should anything be added, removed, or renamed?
• What should the skill be called? (suggest a name based on the dominant theme)

Stage 2: Skill Creation

Once the user confirms, build a full cross-platform skill package following the skill creation pipeline. See references/creation-pipeline.md for the detailed 5-phase process.

Output Structure

<skill-name>-skill/
├── SKILL.md          # Frontmatter + body, starts with "# /<skill-name>"
├── scripts/          # Functional code (no placeholders, no TODOs)
├── references/       # Detailed docs loaded on demand
├── install.sh        # Cross-platform installer (auto-detect platform)
└── README.md         # Installation instructions

SKILL.md Frontmatter Format

---
name: <skill-name>-skill
description: >-
  <activation description with domain keywords, <=1024 chars>
license: MIT
metadata:
  author: <user>
  version: 1.0.0
  created: <today>
  last_reviewed: <today>
  review_interval_days: 90
---

SKILL.md Body Rules

• Starts with # /<skill-name> (the slash invocation trigger)
• Body under 500 lines — move details to references/
• Include a ## Trigger section with invocation examples
• Functional instructions, not documentation to read

Auto-Install

After creation, detect the current platform and install:

• ~/.kiro/ exists → copy to ~/.kiro/skills/<skill-name>-skill/
• ~/.claude/ exists → copy to ~/.claude/skills/<skill-name>-skill/
• ~/.agents/ exists → copy to ~/.agents/skills/<skill-name>-skill/
• Fall back to showing manual install instructions

Tell the user exactly how to invoke the new skill.

Share (Optional)

Ask if they want to share with their team. If yes, init a git repo and create a remote via gh or glab if available.

Quality Gates

Before delivering, verify:

• SKILL.md frontmatter has name, description, license, metadata
• name is lowercase, hyphenated, 1-64 chars
• description is <=1024 chars with domain keywords
• SKILL.md body starts with # /<skill-name>
• Body is under 500 lines
• No hardcoded API keys, credentials, or secrets in any file
• All code is functional (no TODOs, no placeholders, no stubs)
• install.sh is executable and handles platform auto-detection

If any check fails, fix it before delivering.

Keywords for Detection

Actions: create skill, build skill, distill session, capture workflow, package session, save as skill, turn into skill, extract patterns

Contexts: chat session, conversation history, session work, session knowledge, workflow patterns, domain knowledge

Platforms: cross-platform, agent skill, SKILL.md, Kiro, Claude Code, Copilot, Cursor, Gemini CLI, Codex CLI