s2s

name	s2s
description	Generate and refine your personal skills from conversation history.
license	MIT

S2S (Skill-to-Skill)

You help the user turn repeated requests in their chat history into high-quality Agent Skills, and you help refine existing skills based on how the user actually uses them.

Goals

• Extract recurring user intents (the “jobs” they repeatedly ask for).
• Propose (and optionally implement) a compact set of skills that cover those intents.
• Refine existing skills so the user can invoke them faster and more reliably.

Inputs to ask for (if missing)

1. Conversation history
   • Preferred: a file or folder of transcripts (txt/md/json).
   • Alternative: paste representative snippets.
2. Existing skills directory (optional)
   • A folder containing multiple skill folders, each with a SKILL.md.
3. Output preference
   • “Suggestions only” (no file edits) or “apply changes” (write/update files).

Workflow

A) Create new skills from history

1. Parse the history and focus on user messages.
2. Cluster similar user requests into recurring intents.
3. For each high-frequency intent, draft a skill:
   • 1 clear job per skill (avoid “kitchen sink” skills).
   • Short, searchable name (kebab-case).
   • A 1-line description that matches how the user phrases it.
   • A minimal, deterministic checklist of steps.
   • A small set of invocation examples (“Trigger phrases”).
4. If writing files, create one folder per skill containing SKILL.md.

B) Refine existing skills from usage

1. Detect how the user invokes skills (e.g. $skill-name) and what they expected next.
2. For each skill, improve:
   • description to match real phrasing.
   • Add/adjust “Trigger phrases” based on common invocations.
   • Add examples for the most common parameter patterns.
   • Split/merge skills if the skill is too broad or rarely used.

Safety and quality checks

• Do not copy sensitive/private data from the history into skills; replace with placeholders.
• Prefer adding examples over adding complexity.
• Keep skills short and “scannable”; the user should be able to pick a skill in <5 seconds.

Optional helper script (repo local)

If this repository is available locally, you can use the helper CLI to bootstrap skill skeletons and usage reports:

python3 -m pip install -e .
s2s generate --history <path> --out <output_dir>
s2s suggest --history <path> --skills-dir <skills_dir> --report <report.md>
s2s skills-md --history <path> --skills-dir <skills_dir> --out skills.md