AgentKit
DocsKitsCLI ReferenceDesktop App
BetaYou are reading docs for the beta channel (2.13.0-beta.20). Features may change before the next stable release.Switch to stable →

Skills

Manage Marketing Skills with ak:skill

Route a project-scoped Skill request through create, add, fix-logs, optimize, plan, or update while keeping research, file mutation, and user-scope changes controlled.

Use ak:skill to create or improve a reusable Marketing Skill in the current project. It routes one of six actions to a focused workflow, keeps SKILL.md concise through progressive disclosure, and distinguishes review-first planning from actions that edit files directly.

Choose ak:skill for a Skill lifecycle task

Use ak:skill when

  • You want to create a Marketing Skill from a description, documentation URL, or repository source.
  • You need to add references or scripts to an existing Skill.
  • You want an optimization plan and review gate before changing a Skill.
  • You need to fix a Skill from a project-root logs.txt file.
  • You want to plan a new Skill or apply a bounded content update.

Choose another workflow when

  • You still need to decide between a Skill, Agent, command, and workflow. Use ak:kit-builder for component selection and scaffolding.
  • You need to install or refresh Marketing Kit rather than author project content. Follow Installing kits.
  • You only need to read or explain a Skill. Ask for analysis without selecting a mutating action.
  • You intend to change a user-scope Skill. That requires separate, explicit authorization; project scope is the default.

Prepare the target and evidence

Before you start:

  • Confirm Marketing Kit is installed for the current runtime and scope.
  • Read repository instructions and locate the project Skill directory.
  • Preserve unrelated changes and name the target Skill precisely.
  • For fix-logs, inspect the project-root logs.txt and remove secrets or personal data before using it.
  • For a URL or repository source, define the allowed network, download, and external-tool boundary before research begins.
RuntimeInvocationAvailability boundary
Claude Code/ak:skill ...Runs the Marketing-owned lifecycle workflow with its Claude-oriented authoring references.
Cursor/ak:skill ...Uses the projected Skill through slash invocation; named tools or subagents still depend on the active runtime.
Codex$ak:skill ...Uses native Skill discovery; Claude-specific paths or capabilities are guidance, not automatic Codex equivalents.

Run the Skill

The first positional word is the action. This example plans a new Skill and keeps implementation behind a later approval:

/ak:skill plan campaign-brief "Create a project-scoped Skill that turns an approved launch brief into reviewable draft assets; do not implement yet"

Choose an action deliberately

ActionRequired inputMutation and review boundary
add <skill-name> <ref-prompt>Existing Skill and the reference or script to addEdits the project Skill after resolving missing details.
create <prompt-or-url>Description, documentation URL, or repository sourceResearches as needed, then creates a project Skill; network and tool installation need approval.
fix-logs [skill-name]Project-root logs.txt plus an optional targetDiagnoses the logged failure, then edits only the authorized project Skill.
optimize <skill-name> [prompt]Existing Skill and optional constraintsProposes a plan, asks for review, writes the approved plan, then asks again before implementation.
plan <skill-name> [prompt]New Skill name and desired outcomeProduces a review-first implementation plan; approval to save a plan is not approval to implement it.
update <skill-name> [prompt]Existing Skill and requested changeApplies a direct bounded update; use optimize when you need a plan gate first.

Only these six actions are routed by the public Skill contract. Do not rely on an unlisted internal reference as another supported action.

Understand the stages

  1. Parse the action and target. The Skill asks for missing material rather than guessing a Skill name or scope.
  2. Load the matching reference. Each action has its own rules for research, review, output, and mutation.
  3. Inspect project scope. Existing SKILL.md, references, scripts, repository instructions, and relevant logs are read before changes.
  4. Research only as authorized. Documentation lookup or repository analysis may use network access and external tools; broad crawling is not implied by a simple local update.
  5. Plan or edit according to the action. plan and optimize insert review gates. add, create, fix-logs, and update can proceed to authorized project-file changes.
  6. Validate progressive disclosure. The entry file stays concise while detailed guidance and deterministic helpers move to focused resources.
  7. Report evidence and remaining authority. The run lists paths, checks, unresolved risks, and any Git or distribution step not yet approved.

Keep research and mutation controlled

Project scope is the default authority

Do not edit a user-scope Skill unless the user explicitly authorizes that scope and exact target. A URL or GitHub source may trigger network access, repository summarization, or a request to install a helper; approve those effects separately. Never place credentials, private source, or personal data in Skill files, plans, logs, or reports.

Planning approval has two boundaries: approving a proposal allows the workflow to save the specified plan; implementation starts only after the separate implementation confirmation. A later commit, push, package export, publication, or release also remains a separate decision.

Verify outputs and evidence

A complete run should provide:

  • The selected action, target Skill, project scope, and accepted non-goals.
  • Exact files created or changed, including new references or scripts.
  • A concise SKILL.md whose description states when the Skill should activate.
  • Evidence that referenced files exist and scripts have representative tests or a documented reason they were not run.
  • For plan or optimize, the approved plan location, phase files, success criteria, risks, and implementation status.
  • A final diff summary with unresolved questions and all unapproved external or distribution actions.

Troubleshoot and respect limits

SymptomSafe next step
The action or arguments are missingSupply one supported action and a precise project Skill target.
logs.txt contains secrets or unrelated failuresStop, redact sensitive data, and narrow the evidence before fix-logs.
URL research grows beyond the targetRestate allowed domains, links, files, and network/tool permissions.
The run starts editing during plan or before optimization approvalStop and return to the review gate; plan approval and implementation approval are separate.
The Skill is not discovered after editingConfirm path, runtime, scope, and session refresh with Runtime cannot find a Skill or Agent.

The bundled action references are primarily Claude-oriented and may name tools, Agents, or paths unavailable in another runtime. The workflow cannot prove an external source is complete, install tools without approval, or make a project Skill part of the released Marketing package automatically. Continue with Marketing Kit or use ak:kit-builder for a non-Skill component.