Skill Optimizer

中文版

Three Agent Skills for turning coding-agent work into better SKILL.md files:

• skill-miner — mine coding-agent history, archives, memories, and repeated work to surface skill-worthy workflows with evidence.
• skill-personalizer — audit and adapt newly created, downloaded, forked, or community skills to one user's own tools, habits, directories, and session history.
• skill-generalizer — turn local, private, personal skills into publishable skills for GitHub, marketplaces, teams, or public sharing.

Current release: v2.0.0. This is a major redesign from the original single-skill optimizer.

Project site: https://hqhq1025.github.io/skill-optimizer/

The split is intentional. Generating, personalizing, and publishing skills are different jobs:

See Research Background for related agent-skill ecosystems, comparable projects, and papers that motivate session mining, skill libraries, trigger auditing, progressive disclosure, and lifecycle governance.

Installation

Copy the command below into your agent's chat:

Claude Code

Install the skills from https://github.com/hqhq1025/skill-optimizer

Codex

Install the skills from https://github.com/hqhq1025/skill-optimizer into ~/.codex/skills/

Other Agent Skills-compatible agents

Install the skills from https://github.com/hqhq1025/skill-optimizer into ~/.agents/skills/

Manual install:

git clone https://github.com/hqhq1025/skill-optimizer.git /tmp/skill-optimizer
mkdir -p ~/.agents/skills
cp -r /tmp/skill-optimizer/skills/skill-miner ~/.agents/skills/
cp -r /tmp/skill-optimizer/skills/skill-personalizer ~/.agents/skills/
cp -r /tmp/skill-optimizer/skills/skill-generalizer ~/.agents/skills/
rm -rf /tmp/skill-optimizer

For Codex-only installs, use ~/.codex/skills/:

git clone https://github.com/hqhq1025/skill-optimizer.git /tmp/skill-optimizer
mkdir -p ~/.codex/skills
cp -r /tmp/skill-optimizer/skills/skill-generalizer ~/.codex/skills/
cp -r /tmp/skill-optimizer/skills/skill-miner ~/.codex/skills/
cp -r /tmp/skill-optimizer/skills/skill-personalizer ~/.codex/skills/
rm -rf /tmp/skill-optimizer

For Claude Code-only installs, use ~/.claude/skills/.

Platform Support

The safest public layout is skills/<name>/SKILL.md in the repo plus install instructions that copy into .agents/skills/ or the target agent's native skill directory.

Usage

Ask for the direction you want:

Mine my coding-agent history and find repeated workflows that should become skills.

Audit and tune my installed skills; tell me which ones are undertriggering, too noisy, or too generic.

Turn this local skill into a public GitHub-ready skill.

I downloaded this skill. Tune it to my local workflow and usage habits.

This skill does not trigger when I say things naturally. Personalize it for me.

What Each Skill Does

skill-miner

• coding-agent session history, memory summaries, repo notes, repeated scripts, and project folders
• recurring user intents, shorthand, tool chains, artifacts, and verification patterns
• candidates that are repeated and non-obvious enough to become skills
• whether a candidate should stay personal, be generalized for publication, or be skipped
• includes scripts/scan_sessions.py for a deterministic first-pass scan of Codex, Claude Code, Gemini/Antigravity task files, and exported transcripts from other agents
• includes archived Codex sessions and rollout summaries by default, with flags to disable archive/summary sources

Example:

python3 skills/skill-miner/scripts/scan_sessions.py --days 30 --limit 300 --min-count 3
python3 skills/skill-miner/scripts/scan_sessions.py --export ~/Downloads/cursor-chat-export.json
python3 skills/skill-miner/scripts/scan_sessions.py --patterns ./my-patterns.json
python3 skills/skill-miner/scripts/scan_sessions.py --no-include-archives --no-include-summaries

skill-generalizer

• private paths, hosts, credentials, account names, transcript quotes, and internal repo facts
• public portability of commands, examples, README claims, and install instructions
• frontmatter that describes when to use the skill rather than the workflow
• packaging structure for public distribution

skill-personalizer

• local installed copies and nearby project instructions
• real user phrasing and recurring task patterns
• preferred CLIs, MCP tools, paths, aliases, and verification commands
• undertrigger, overtrigger, and unnecessary-question friction
• original optimizer-style audit checks: trigger fit, user reaction, workflow completion, static quality, conflicts, environment consistency, token economics, and P0/P1/P2 fixes

Compatibility

Works with agents that support the Agent Skills folder convention:

• Claude Code
• Codex
• Cursor
• OpenCode
• Gemini CLI

Research Background

This project is informed by Agent Skills ecosystem work and LLM-agent research on externalized memory, skill libraries, retrieval/routing, and long-context behavior. See docs/research-background.md.

AI And Search Visibility

• Project site: https://hqhq1025.github.io/skill-optimizer/
• LLM summary: llms.txt
• Full LLM context: llms-full.txt
• Structured metadata: repo-metadata.json

License

MIT