# Add to your Claude Code skills
git clone https://github.com/jturntdev/kryptonkrypton is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by jturntdev. Stop AI agents from shipping plausible wrong work. It has 73 GitHub stars.
krypton's catalog security scan is still queued. You can run an instant dependency and prompt-injection check now with the "Scan for vulnerabilities" button above.
Clone the repository with "git clone https://github.com/jturntdev/krypton" and add it to your Claude Code skills directory (see the Installation section above).
krypton is primarily written in Shell. It is open-source under jturntdev on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh krypton against similar tools.
No comments yet. Be the first to share your thoughts!
Unlocks once the catalog security scan passes (runs nightly).
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.
Current version: 0.1.3
Stop AI agents from shipping plausible wrong work.
Krypton is an operator discipline for running Claude Code, Codex, and other AI coding agents against production-grade codebases.
The failure mode Krypton is built for is not bad syntax. It is the right-looking feature that compiles while living on the wrong layer, inventing a second source of truth, skipping cutover, or claiming success without proof from the real path.
Krypton forces the plan before the code:
Built for /goals, Claude Code skills/plugins, Codex skills, and operators who
need agent speed without letting the codebase rot.
Modern agents can create weeks of architectural debt in one enthusiastic session. Krypton turns "build this" into an operational contract before the agent touches code:
That contract is what lets an operator keep using agents on production systems without letting the codebase become a pile of current-looking alternatives.
/goalsKrypton is best used with /goals or any equivalent goal-based agent workflow.
It is not meant to be a one-shot "please implement this" prompt.
The planning session creates the durable handoff:
docs/goals/<goal-slug>/PLAN.md
docs/goals/<goal-slug>/GOAL.md
PLAN.md is the source plan. GOAL.md is the compact /goal prompt. Start the
execution session with that prompt, then load krypton-execution so the main
agent preserves the plan's ownership, cutover, review, and evidence gates.
If your harness does not have /goals, paste the contents of GOAL.md into a
fresh Codex or Claude session. The shape still works.
Use the plugin route when you want Claude Code to manage Krypton as a plugin:
/plugin marketplace add jturntdev/krypton
/plugin install krypton@krypton-dev
/reload-plugins
Claude Code namespaces plugin skills, so invoke them as:
/krypton:krypton-planning
/krypton:krypton-execution
This repo includes the Claude plugin files Claude Code expects:
.claude-plugin/plugin.json
.claude-plugin/marketplace.json
skills/krypton-planning/SKILL.md
skills/krypton-execution/SKILL.md
Use manual install when you want the skills available as personal Claude Code skills without the plugin marketplace:
git clone https://github.com/jturntdev/krypton.git
mkdir -p ~/.claude/skills
cp -R krypton/skills/krypton-planning ~/.claude/skills/
cp -R krypton/skills/krypton-execution ~/.claude/skills/
Manual personal skills invoke without a plugin namespace:
/krypton-planning
/krypton-execution
For Codex skills:
git clone https://github.com/jturntdev/krypton.git
mkdir -p ~/.codex/skills
cp -R krypton/skills/* ~/.codex/skills/
krypton-planning: turn a request into an outcome contract, architecture
slice, task plan, evidence gate, and /goal handoff prompt.krypton-execution: execute an approved plan without drifting from ownership,
cutover, or proof requirements.krypton-planning with a feature request, bug, refactor, migration, or
architecture goal.docs/goals/<goal-slug>/PLAN.md
docs/goals/<goal-slug>/GOAL.md
PLAN.md is the full implementation plan. GOAL.md is the short /goal
prompt or handoff prompt for the next session.GOAL.md prompt in Codex or Claude.krypton-execution when you want the main agent to use the same
ownership, cutover, review, and evidence discipline.The workflow is intentionally two-stage:
rough request
-> krypton-planning
-> PLAN.md + GOAL.md
-> /goal handoff
-> krypton-execution
-> main-agent implementation + review gates + acceptance evidence
Krypton works best when the harness supports named agents:
explorer: read-only source-of-truth and architecture mappingplan-reviewer: PRE and POST alignment checksreviewer: runtime correctness, security, and evidence checkmaintainer: codebase-shape, duplication, and cutover-debt checkverifier: focused proof from the real pathSee docs/required-roles.md for the role expectations. If your harness does
not support named agents, the main agent can still follow the same gates, but
independent exploration and review will be weaker.
Krypton keeps prompt templates individual instead of centralizing them into one large prompt file. Each skill owns the prompts it actually uses:
skills/krypton-planning/plan-reviewer-prompt.mdskills/krypton-execution/post-plan-reviewer-prompt.mdskills/krypton-execution/reviewer-prompt.mdskills/krypton-execution/maintainer-prompt.mdBad agent path:
"Add sentiment to the dashboard."
The agent adds frontend-only sentiment logic and calls it done.
Krypton path:
Truth owner: market intelligence producer
Contract boundary: typed sentiment payload
Cutover: dashboard reads producer output only
Evidence: API response plus browser state showing the real payload
See examples/ and tests/pressure-scenarios/ for more.
This is the first public cut. It is intentionally small: two skills, individual prompt templates, agent role expectations, examples, pressure scenarios, and a validation script.
Krypton uses SemVer-style versions while the public package stabilizes.
VERSION.codex-plugin/plugin.json,
.claude-plugin/plugin.json, and .claude-plugin/marketplace.jsonvX.Y.ZVersion policy:
/goal workflow shapeRun:
./scripts/validate.sh
The validator checks required files, skill metadata, JSON plugin metadata, and public-safety issues such as placeholders or private project references.
MIT