What value does it add?
More reliable, composable, and traceable Agent work—especially when a task spans tools, permissions, evidence, and time.
Reliable, composable, and traceable Agent work.
Kungfu Skills · product preview
Make Agent skills reliable enough for real work.
Kungfu Skills keep the simplicity of reusable Agent instructions, then add exact identity, Work fit, declared dependencies, visible effects, proof, and recovery for complex jobs.
The goal is not a smarter prompt. It is work that is easier to repeat, combine, inspect, and continue without mysterious changes or hidden authority.
How is it different?
A normal Skill is a useful playbook. A Kungfu Skill is designed to connect that playbook to exact Work, dependencies, effects, proof, and recovery while authority stays elsewhere.
It binds the playbook to exact Work, effects, proof, and authority boundaries.
Do I need to code?
No. Using one should feel like asking your Agent to do the job. The simplest Skill can start as one SKILL.md file, and an Agent can help author it.
No—ask your Agent in natural language.
The value
Not a smarter Skill. A more dependable way to use one.
Repeat with confidence
Pin the exact instructions and packaged content used for the work, so a later update cannot silently change what happened before.
Combine without hidden power
Declare the integrations and effects a Skill needs while keeping permission and execution decisions in their own governed systems.
Explain the work later
Carry proof requirements, recovery steps, provenance, and predecessor links so people and Agents can reconstruct which Skill influenced a result.
The difference
A familiar idea, with an engineering-grade work contract around it.
| Question | A typical Agent Skill | A Kungfu Skill |
|---|---|---|
| What it starts with | Instructions, often with supporting scripts. | The same approachable instruction layer, rooted as one complete, inspectable package. |
| Which version ran | Often the file as it exists now. | A stable key, immutable revision, complete content root, and predecessor chain. |
| What it may do | Tools and permissions may be implied by prose. | Dependencies and effects are declared; separate systems still decide capability and execution. |
| What counts as done | The Agent may infer success from steps or output. | Proof is explicit, and Work—not the Skill—owns acceptance and completion. |
| What you can explain later | That a named Skill was probably used. | The exact content, provenance, effects, evidence needs, and recovery meaning tied to that Work. |
Example · release a project
The request stays simple. The safety underneath becomes explicit.
The user can still ask in one sentence. A Kungfu Skill is designed to make the hidden coordination inspectable without making the user operate it by hand.
“Release this project.”
Fit the exact Skill to the Work
Bind one immutable Skill revision and complete content root to the release job.
Resolve what the work depends on
Name the required build and repository integrations without letting Skill prose grant them authority.
Make side effects visible
Declare process, filesystem, and external writes before the operational path uses them.
Keep the evidence that matters
Require test, build, release, and user-acceptance evidence instead of equating output with completion.
Authority boundary: The Skill explains and orchestrates how. Work owns selection and completion; KFX owns executable capability; evidence systems preserve what actually happened.
How to use it
Use it in natural language. Review decisions, not implementation plumbing.
Ask for the outcome
Tell your Agent what you want done, as you already do today.
Let the Agent propose a Skill
The intended flow matches an exact Skill to the Work and shows important dependencies or effects.
Review meaningful choices
You approve authority, destructive effects, or acceptance decisions when they matter—not routine code.
Receive a result with proof
The outcome keeps evidence and recovery context so another person or Agent can continue later.
No coding required to use one. You do not need to know how to code to use a Kungfu Skill. Creating the simplest one can begin with a single SKILL.md file, and an Agent can draft that from your description. More advanced integrations may require developer work behind the Skill, but using them should not.
Availability · honest preview boundary
The contract is defined. End-to-end adoption is still being staged.
The protected development line contains the first context loop and the new reliability contract. This preview does not claim that the complete registry, Work selection, automatic composition, or distribution experience is released today.
Technical evidence
Read the contract and exact sources
Identity, Work scope, effects, proof, authority boundaries, architecture, digests, and non-claims.
Product direction
See what is current, next, and later
Staged runtime adoption is separated from future automatic composition and governed distribution.