kk-r/skillify-skill/skillify

Skillify — Turn Any Session Into a Reusable Skill

You are capturing this session's repeatable process as a reusable SKILL.md file that follows the agentskills.io open standard — compatible with Claude Code, Cursor, GitHub Copilot, Gemini CLI, VS Code, and 30+ other agent platforms.

Phase 0: Gather Session Context

You don't have direct access to session memory, so reconstruct it now using three complementary sources.

Step A: Review Conversation History

Look back through the entire conversation. Extract:

• Goal: What did the user ask you to accomplish?
• Steps taken: Ordered list of actions (tools used, files touched, commands run)
• Corrections: Where did the user redirect your approach? These become Rules in the skill.
• Tools & permissions: Which tools were critical? Note permission patterns (e.g., Bash(gh:*) not just Bash)
• Decision points: Where did you or the user choose between alternatives?

Step B: Check Git Artifacts

Gather recent changes to fill gaps in conversation context:

!git diff --stat 2>/dev/null | head -30

!git log --oneline -10 2>/dev/null

Step C: Detect Project Context

Auto-detect the project's tooling so the generated skill uses the right commands:

!{ [ -f package.json ] && echo "NODE: $(cat package.json | grep -E '\"(name|test|build|lint)\"' | head -5)"; [ -f Makefile ] && echo "MAKE: $(head -20 Makefile | grep '^[a-z].*:')"; [ -f Cargo.toml ] && echo "RUST: $(head -5 Cargo.toml)"; [ -f go.mod ] && echo "GO: $(head -3 go.mod)"; [ -f Gemfile ] && echo "RUBY: $(head -5 Gemfile)"; [ -f pyproject.toml ] && echo "PYTHON: $(head -10 pyproject.toml)"; [ -f requirements.txt ] && echo "PYTHON: requirements.txt found"; } 2>/dev/null || echo "No standard project files detected"

Phase 1: Interview the User

Use AskUserQuestion for ALL questions. Never ask questions via plain text. Iterate each round until the user is satisfied. The user always has a freeform "Other" option — do NOT add your own "Needs tweaking" option.

Round 1: High-Level Confirmation

• Present your summary from Phase 0
• Suggest a name (lowercase, hyphens, max 64 chars per agentskills.io spec) and one-line description
• Suggest high-level goal(s) and success criteria
• Ask the user to confirm, rename, or adjust

Round 2: Structure and Scope

• Present steps as a numbered list. Tell the user you'll dig into per-step detail next round.
• If the skill needs arguments, suggest them based on what you observed. Clarify what a future user would provide.
• Ask execution context:
• inline (default) — runs in current conversation, user can steer mid-process
• fork — runs as isolated sub-agent, better for self-contained tasks
• Ask save location:
• This repo (.claude/skills/<name>/SKILL.md) — project-specific workflows
• Personal (~/.claude/skills/<name>/SKILL.md) — follows user across all repos

Round 3: Step-by-Step Detail

For each major step (skip if obvious), ask:

• What does this step produce that later steps need? (artifacts: PR URL, commit SHA, file path)
• What proves this step succeeded? (success criteria — required on every step)
• Should the user confirm before proceeding? (human checkpoint — for irreversible actions)
• Can any steps run in parallel? (concurrent steps use sub-numbers: 3a, 3b)
• What are hard rules? (constraints from user corrections, must/must-not)

Do multiple rounds if there are more than 3 steps or complex decision points.

Round 4: Triggers and Edge Cases

• Confirm when this skill should be invoked — suggest trigger phrases
• Example: "Use when the user says 'cherry-pick', 'hotfix', or 'CP this PR to release'"
• Ask about edge cases, gotchas, or failure modes to handle
• Ask if the skill should be cross-platform (if yes, note platform-specific commands)

Stop interviewing once you have enough. Don't over-ask for simple 2-3 step processes.

Phase 2: Write the SKILL.md

Generate a SKILL.md following the agentskills.io standard.

Frontmatter Template

---
name: {{skill-name}}
description: >
  {{One-line description. Start with an action verb. Under 1024 chars.
  Include "Use when..." trigger context so agents know when to activate.}}
license: MIT
metadata:
  author: {{user or org name}}
  version: "1.0.0"
allowed-tools:
  {{Minimum permission patterns observed. Use Bash(gh:*) not Bash.}}
---

Body Template

# {{Skill Title}}
 
{{Brief description of what this skill does and its goal.}}
 
## Inputs
 
-`$arg_name`: Description of this input
 
## Goal
 
{{Clearly stated goal. Include concrete success artifacts
(e.g., "an open PR with CI passing" not just "code changes").}}
 
## Steps
 
### 1. {{Step Name}}
 
{{Specific, actionable instructions. Include commands where appropriate.}}
 
**Success criteria**: {{How to know this step is done.}}
 
### 2. {{Step Name}}
 
...

Writing Rules

Frontmatter:

• name: lowercase, hyphens only, max 64 chars, must match directory name
• description: under 1024 chars, start with action verb, include "Use when..." triggers
• allowed-tools: minimum permissions needed — use patterns like Bash(gh:*) not Bash

Body:

• Success criteria on EVERY step — this is required, not optional
• Use per-step annotations where helpful:
• Execution: Direct (default), Task agent, Teammate (parallel), [human]
• Artifacts: Data this step produces for later steps
• Human checkpoint: Pause for user confirmation (irreversible actions)
• Rules: Hard constraints (especially from user corrections during original session)
• Concurrent steps use sub-numbers: 3a, 3b
• Steps requiring user action get [human] in the title
• Keep simple skills simple — a 2-step skill doesn't need every annotation
• Put large reference material in a references/ subdirectory, not inline

Cross-platform compatibility:

• The agentskills.io standard works across 30+ tools
• Avoid Claude Code-specific frontmatter fields when possible
• Use standard fields: name, description, license, metadata, allowed-tools
• Tool-specific fields (like when_to_use, context, arguments) are fine — agents that don't understand them simply ignore them

Phase 3: Review and Save

1. Output the complete SKILL.md as a yaml code block so the user can review with syntax highlighting
2. Ask for confirmation via AskUserQuestion: "Does this SKILL.md look good to save?"
3. On approval:

• Create the skill directory
• Write the SKILL.md file
• If the skill has reference files, create a references/ subdirectory

1. Confirm to the user:

• Where the skill was saved
• How to invoke: /{{skill-name}} [arguments]
• That they can edit the SKILL.md directly to refine it
• That the skill follows agentskills.io and works across compatible agent platforms
• Remind them to restart Claude Code (skills are loaded at startup)