A provider-agnostic agentic engineering stack for any project, in two parts:
- Skills. Curated Agent Skills and a shared working style for Claude Code, Cursor, Codex CLI, and Antigravity CLI.
- Factory. A controller that runs one coding task on a throwaway, isolated EC2 worker and ends with a pull request.
npx skills@latest add olibyte/agent-toolkit \
-a claude-code -a cursor -a codex -a antigravity -a antigravity-cli -yInstall into the project. A local clone uses the clone path in place of olibyte/agent-toolkit.
Copy templates/AGENTS.md and templates/CLAUDE.md to the repo root when those files are missing. Copy templates/models.md to .agents/models.md when you want to pin subagent models.
Commit .agents/skills/, .claude/skills/, skills-lock.json, AGENTS.md, and CLAUDE.md. Leave .agents/models.md uncommitted when the team does not share models. Run handoff before switching tools.
Working style from pstack: match a playbook, name the data shape, delegate the implementation, check the real artifact. Casual turns skip it.
Codex: $poteto-mode. Claude Code, Cursor, and Antigravity: /poteto-mode. Playbooks live in skills/poteto-mode/SKILL.md. /poteto-help picks a skill or playbook. /ketchup catches you up on this chat since your last message.
These live under optional/ so a default Codex session stays small. Drop any -s you do not need.
npx skills@latest add olibyte/agent-toolkit --full-depth \
-s swarm -s interrogate -s figure-it-out \
-s create-verification-skill -s maintain-verification-skill \
-s reflect -s automate-me \
-a claude-code -a cursor -a codex -a antigravity -a antigravity-cli -yInvoke as $name or /name.
- swarm. Fan out workers, wait, return one report.
- interrogate. Several models review one diff and return a verdict. The diff stays as it is.
- figure-it-out. Write a playbook when none fits, then run it.
- create-verification-skill. Write a project-local
verify-<app>skill under.agents/skills/. - maintain-verification-skill. Recheck that map against source and a live pass, then open at most one PR of proven corrections.
- reflect. Mine this chat for durable lessons and route each one to a skill edit.
- automate-me. Draft a personal mode skill from recent transcripts.
poteto-mode calls swarm, interrogate, figure-it-out, and reflect when they are installed, and skips them with a reason when they are absent.
factory/ runs one coding task on a throwaway worker and ends with a pull request. You write a small task file with the repo, the change (a shell command, or a prompt for Claude Code or Codex), and the checks. The factory clones the repo, makes the change, runs the checks, pushes a branch, and opens the PR. It posts each step to Slack, then deletes the worker.
node factory/cli.ts run factory/examples/proof.task.json --worker local # or --worker ec2The local worker needs Node 22.18 or later and gh. The EC2 worker also needs an AWS account and Terraform. factory/README.md covers setup. The factory is separate from the skills, and npx skills add does not install it.
Kun Chen's quota-axi reports local subscription quota windows (Claude, Codex, Cursor, Copilot, and others) so the agent can see remaining runway before it spends more. The report is data only.
npx skills@latest add kunchenguid/quota-axi --skill quota-axi -g \
-a claude-code -a cursor -a codex -a antigravity -a antigravity-cli-g installs the skill for every project. Omit it to keep the skill in the current project.
One command per source. Leave off -y so you can skip skills. Same -a flags as above.
- mattpocock/skills. Planning, grilling, specs, and issue trackers. Skip
handoff,tdd, andteach(this catalog already ships them), then runsetup-matt-pocock-skillsonce. - vercel-labs/agent-skills. React and Next.js performance rules and web design audits.
- vercel/next.js. Next.js workflows such as
next-dev-loop. Install only in a Next.js app. Next.js 16.3 and later already writes framework docs intoAGENTS.md.
npx skills@latest add <owner/repo> \
-a claude-code -a cursor -a codex -a antigravity -a antigravity-cli