# Skills preview surface 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. Status: site-synthesis Maturity: protected-source-alpha-preview Canonical human entry: https://libkungfu.dev/skills/ Machine entries: - https://libkungfu.dev/skills/manifest.json - https://libkungfu.dev/skills/architecture.json - https://libkungfu.dev/skills/capability-map.json - https://libkungfu.dev/skills/llms.txt Reader paths: - decide / Decide whether the work deserves a Skill: Look for repeatable delegated work with stable inputs, a bounded outcome, and evidence worth retaining. Sources: skills-architecture, skills-adr - understand / Understand the current source contract: Separate Skill source, compact catalog, context envelope, Agent consumption, and the KFX trust gate. Sources: skills-contract, skills-architecture - audit / Audit what this preview does not prove: Read the partial implementation status, known limits, future-picture labels, and explicit non-claims before relying on the surface. Sources: skills-adr, skills-known-limits Verified current protected-source facts: - exact-identity / Every definition has an exact identity [upstream-fact; accepted-contract]: The accepted contract binds a stable key, immutable revision, complete content root, provenance, and predecessor coordinate so later changes cannot reinterpret old bytes. Sources: skills-contract, skills-adr - work-boundary / Work owns selection and completion [upstream-fact; accepted-contract]: A definition declares its Work scope and binding policy, but catalog visibility, instruction loading, invocation, process exit, or Agent output cannot select or complete Work. Sources: skills-contract, skills-adr - declared-effects / Dependencies and effects must be declared [upstream-fact; accepted-contract]: Operational and domain definitions name exact KFX or Profile dependencies and bind every effect to a declared authority; prose cannot grant capability or copy another system's authority. Sources: skills-contract, skills-architecture, skills-adr - proof-recovery / Proof and recovery are part of the contract [upstream-fact; accepted-contract]: Every definition carries explicit proof requirements, completion ownership, recovery steps, runtime compatibility, and historical preservation rules. Sources: skills-contract, skills-adr - partial-status / The contract is accepted; adoption is staged [upstream-fact; accepted-partial]: The deterministic definition contract and rejection fixtures are retained on dev, while registry, lifecycle, Work selection, and end-to-end runtime adoption remain follow-up work. Existing context-loop slices remain readable during migration. Sources: skills-contract, skills-architecture, skills-adr, skills-known-limits Agent suggestion and authoring guidance: - suggest-repeatable / Suggest a Skill for repeated or complex delegated work [suggest; site-synthesis]: Suggest one when the same intent, dependencies, effects, checks, and recovery needs recur—or when one long-running task must survive handoffs and change. Sources: skills-architecture, skills-adr - create-bounded / Start with instructions, not code [create; site-synthesis]: The smallest Skill is one SKILL.md. Add structure, exact dependencies, effects, proof, and recovery only when the work needs them. Sources: skills-architecture, skills-adr - avoid-one-off / Keep one-off work in the conversation [avoid; site-synthesis]: Do not create a Skill merely to preserve a transient prompt, a single workaround, or prose that has no reusable verification boundary. Sources: skills-architecture - avoid-authority / Do not hide permissions inside instructions [avoid; non-claim]: A Skill may request or explain an action; it must not imply that its text grants capabilities, installs runtime artifacts, or bypasses KFX admission. Sources: skills-adr, skills-known-limits Architecture nodes: - delegated-work / Repeatable delegated work [site-synthesis; site-synthesis; reader-guidance]: A bounded recurring job whose intent, checks, evidence, and recovery deserve a reusable form. Sources: skills-architecture, skills-adr - skill-source / SKILL.md source [current-source; upstream-fact; protected-source-alpha]: The minimum instruction-only source, optionally enhanced with structured metadata. Sources: skills-contract, skills-architecture - skill-catalog / Compact catalog [current-source; upstream-fact; protected-source-alpha]: A small index intended to help an Agent discover relevant instructions without loading every source in full. Sources: skills-contract, skills-architecture - context-envelope / Context envelope [current-source; upstream-fact; protected-source-alpha]: The source-defined catalog, on-demand read declaration, and audit identifiers for a managed Agent context. Sources: skills-contract, skills-architecture - agent-consumer / Agent consumer [current-source; upstream-fact; source-contract-only]: Consumes advertised metadata and may request full instructions; this page does not claim a released runtime path. Sources: skills-architecture, skills-known-limits - kfx-trust-gate / KFX trust and capability gate [current-source; upstream-fact; accepted-architecture]: Retains runtime admission and capability authority below the Skill instruction layer. Sources: skills-architecture, skills-adr - accountability-history / Durable provenance and work history [future-picture; future-picture; future-picture-only]: A future view could connect Skill influence to retained work evidence without making the Skill itself authority. Sources: skills-architecture, skills-adr - governed-distribution / Governed discovery and distribution [future-picture; future-picture; future-picture-only]: Discovery, remote publishing, and dependency acquisition remain future product work with separate trust and release gates. Sources: skills-known-limits Architecture relationships: - recognize: delegated-work may justify authoring skill-source [site-synthesis; site-synthesis] Sources: skills-architecture, skills-adr - derive: skill-source derives skill-catalog [current-source; upstream-fact] Sources: skills-contract, skills-architecture - wrap: skill-catalog is wrapped by context-envelope [current-source; upstream-fact] Sources: skills-contract, skills-architecture - advertise: context-envelope advertises bounded context to agent-consumer [current-source; upstream-fact] Sources: skills-architecture - constrain: kfx-trust-gate keeps runtime authority outside agent-consumer [current-source; upstream-fact] Sources: skills-architecture, skills-adr - retain: agent-consumer could retain influence in accountability-history [future-picture; future-picture] Sources: skills-architecture, skills-adr - distribute: skill-source could enter governed-distribution [future-picture; future-picture] Sources: skills-known-limits Capability map: ## Understand Inspect the current source model without upgrading it into a release claim. - separate-objects / Separate source, catalog, context, Agent, and KFX [defined-in-protected-source; protected-source-alpha; upstream-fact]: Read each object as a different responsibility and keep KFX as the runtime trust boundary. Sources: skills-contract, skills-architecture, skills-adr ## Suggest Use Site guidance to recognize work that benefits from a reusable form. - recognize-repetition / Recognize repeated bounded work [site-guidance; site-synthesis; site-synthesis]: Suggest a Skill when triggers, outcomes, verification, and recovery recur across Agent sessions. Sources: skills-architecture, skills-adr - reject-one-off / Reject one-off prompt capture [site-guidance; site-synthesis; site-synthesis]: Keep transient prompts and unverified workarounds out of a durable Skill surface. Sources: skills-architecture ## Author Start from the smallest source shape and state its boundaries. - minimal-source / Begin with SKILL.md [defined-in-protected-source; protected-source-alpha; upstream-fact]: The protected architecture defines SKILL.md-only instruction source as the minimal form. Sources: skills-contract, skills-architecture, skills-adr - bound-outcome / Name outcome, checks, and recovery [site-guidance; site-synthesis; site-synthesis]: Author for accountable reuse, not for hidden automation or permission transfer. Sources: skills-architecture, skills-adr ## Govern Keep instruction influence inspectable and authority elsewhere. - no-permission-laundering / Preserve the KFX trust gate [accepted-architecture; accepted-partial; upstream-fact]: Skill composition cannot elevate a KFX package or make a refused runtime facet executable. Sources: skills-adr, skills-known-limits - source-digests / Inspect exact sources and digests [available-on-this-preview; site-evidence; site-synthesis]: This preview exposes the exact protected commit, file paths, and SHA-256 digests used for synthesis. Sources: skills-contract, skills-architecture, skills-adr, skills-known-limits ## Future picture Keep useful product directions visible without calling them current capability. - durable-influence / Retain Skill influence with work evidence [future-picture-only; future-picture; future-picture]: A future product could connect selection and instruction loading to durable provenance and recovery evidence. Sources: skills-architecture, skills-adr - governed-discovery / Add governed discovery and distribution [future-picture-only; future-picture; future-picture]: Marketplace, remote publishing, and unresolved dependency acquisition require separate product and release work. Sources: skills-known-limits Immutable sources: - skills-contract [protected-source-contract]: https://github.com/kungfu-systems/kungfu/blob/b4a10ea2a165f0d8eb1010b6d913069b1f17f39f/framework/skill/kungfu-skill.contract.json sha256:9687a533493aa1e6ca2407f633f4ec6039eb0b857afb1db324fefb3c64f3ed01 - skills-architecture [protected-source-architecture]: https://github.com/kungfu-systems/kungfu/blob/b4a10ea2a165f0d8eb1010b6d913069b1f17f39f/docs/architecture/skills.md sha256:4bae5194be6374830fec387be61010e4a8e132330739d13dcf7940a040651663 - skills-adr [accepted-partial-decision]: https://github.com/kungfu-systems/kungfu/blob/b4a10ea2a165f0d8eb1010b6d913069b1f17f39f/docs/adr/KF-ADR-019fee22-e71d-7da9-8a44-9403c21a5d62.md sha256:b439ed0b749c6e1b6f4527356454f0d5aa2024f46a8fe2d5369171ca30a159ce - skills-known-limits [protected-source-known-limits]: https://github.com/kungfu-systems/kungfu/blob/b4a10ea2a165f0d8eb1010b6d913069b1f17f39f/docs/qualification/known-limits.md sha256:b8a913ad020d089842b6d8550025665e307bfd92eb5f5bbb1049d18d7a7b0c47 Source boundary: This page is a Site-owned synthesis of exact files on Kungfu's protected development line. Kungfu owns the Skill contract, implementation status, commands, trust boundary, and qualification claims. The Site owns only reader guidance, composition, and routing. Preview boundary: This additive Site preview does not publish or activate a Kungfu Skill runtime, package, catalog, registry, Agent Brief, or production surface. Future boundary: Future-picture nodes describe a useful direction for accountable authoring, provenance, and distribution. They are not present-tense product or release claims. Non-claims: - Protected source does not prove that a Skill catalog, install flow, selection path, invocation path, or qualification result is shipped in a released Kungfu product. - This preview does not install, enable, select, load, invoke, qualify, certify, publish, or distribute any Skill. - A SKILL.md file is instruction content, not a capability grant, Warrant, trust decision, KFX admission, or permission elevation. - The Site does not claim marketplace discovery, remote publishing, automatic permission elevation, unresolved KFX acquisition, or third-party runtime execution through Skill composition. - Future-picture material is design guidance only and is not evidence for the active Kungfu Skill implementation family.