Quick Ref / Command

Engineer Kit

/ak:skill-creator

Skill Creator

Creates or updates Claude skills through intent capture, research, progressive-disclosure planning, scaffolding, SKILL.md/resources, evals, description optimization, packaging, and optional Kongming advice.

01

Capture intent

02

Research patterns

03

Plan resources

04

Initialize structure

Rule 01

Skills are practical instructions, not documentation

Rule 02

Metadata triggers first; details load progressively

Rule 03

Validation and evals drive iteration

Rule 04

Keep SKILL.md below 300 lines and references focused

Execution Map

Core lanes from input to output, aligned with the command's real execution path.

01

Start

Input, scope, route

  1. 1 Capture intent Use human-in-the-loop clarification to define what the skill does, when it should trigger, and what it should output.
  2. 2 Research patterns Use docs-seeker and research guidance to ground skill best practices before drafting.
  3. 3 Plan resources Choose which instructions stay in SKILL.md and which details belong in scripts, references, agents, or assets.
02

Work

Agent / skill execution

  1. 4 Initialize structure Create the skill folder with required SKILL.md and optional scripts, references, agents, and assets.
  2. 5 Write instructions Keep instructions imperative, concise, pushy in metadata, and free of duplicated material across references.
  3. 6 Test and evaluate Run validation and evals, then compare baseline behavior with and without the skill.
03

Verify

Gate, review, validation

  1. 7 Optimize description Refine pushy metadata, standard terminology, examples, and scope language for trigger accuracy and benchmark scoring.
  2. 8 Package Validate metadata, token use, script dependencies, structure, portability, and marketplace readiness before producing a zip.
  3. 9 Iterate from feedback Generalize from failures or user feedback instead of adding one-off workaround instructions.

Syntax, arguments, subcommands

How to invoke the skill: syntax, positional arguments, shared options, then each subcommand with its own syntax and outcome.

Syntax

/ak:skill-creator <create|update|audit|optimize> [skill-name|path|kit|--all] [--kit <kit>|--project|--user] [--long-horizon] [--apply] [--from-audit <report>] [--advice]

Arguments

<create|update|audit|optimize> Required

Workflow

Required skill argument: create a new skill, update an existing one, audit routing/behavior, or optimize description and evals. Not an ak CLI subcommand.

/ak:skill-creator create release-notes
[skill-name|path|kit|--all]

Target

Skill name, path, kit, or --all for audit/optimize across a kit. Combine with --kit, --project, or --user to set install scope.

/ak:skill-creator audit --all --kit engineer

Shared options

--kit

Kit scope

Target a named kit instead of inferring from the current project.

/ak:skill-creator audit --all --kit engineer
--project

Project scope

Create or update the skill in the current project, not user-global.

/ak:skill-creator create release-notes --project
--user

User scope

Author into user-global skill storage only when the user explicitly asks for that scope.

/ak:skill-creator create release-notes --user
--long-horizon

Long-horizon create

Use the long-horizon creation workflow for skills that must stay durable across many sessions.

/ak:skill-creator create release-notes --project --long-horizon
--apply

Write accepted edits

Write the accepted update or optimize edits. Without --apply, report the proposed patch only.

/ak:skill-creator update release-notes --apply
--from-audit

From audit report

Feed an audit report path into update so the patch stays scoped to diagnosed issues.

/ak:skill-creator update release-notes --from-audit plans/reports/skill-audit.md --apply
--advice

Advisory supervision

Requests Kongming advisory review after planning, after the SKILL.md draft and evaluation results, before packaging or distribution, or when repeated failures block progress. It does not transfer decisions or bypass validation and security checks.

/ak:skill-creator create release-notes --project --advice

Subcommands

create Create
/ak:skill-creator create [skill-name or description] [--kit <kit>|--project|--user] [--long-horizon] [--advice]

Capture intent, author SKILL.md and resources, evaluate, then package only when requested.

Outcome:

/ak:skill-creator create release-notes --project
update Update
/ak:skill-creator update [skill-name|path] [--apply] [--from-audit <report>] [--advice]

Change an existing skill from a request or an audit report. --apply writes accepted edits.

Outcome:

/ak:skill-creator update release-notes --apply
audit Audit
/ak:skill-creator audit [skill-name|path|kit|--all] [--kit <kit>] [--advice]

Diagnose routing, description, and behavior, using measured usage from local sessions without printing their content. Does not rewrite the skill unless you follow with update --apply.

Outcome:

/ak:skill-creator audit release-notes
optimize Optimize
/ak:skill-creator optimize [skill-name|path|--all] [--apply] [--advice]

Tighten trigger metadata, evals, and progressive disclosure. --apply writes the accepted metadata changes.

Outcome:

/ak:skill-creator optimize release-notes --apply

Sample Prompt

Concrete invocations for each flag, subcommand, mode, or route available in the live workflow.

Create a project skill Recommended
/ak:skill-creator create release-notes --project
Use when:
A repeatable workflow should become a versioned skill in the current project.
Expected:
Captures tasks, triggers, outputs, scope, refusal behavior, and test cases; scaffolds the project skill directory, writes concise SKILL.md/resources, validates structure, and packages only when requested.
Update from an audit
/ak:skill-creator update release-notes --from-audit plans/reports/skill-audit.md --apply
Use when:
An audit report already diagnosed routing or instruction issues that should be written back.
Expected:
Reads the audit report, patches only diagnosed issues, writes accepted edits because --apply is set, and leaves unrelated skill files unchanged.
Audit one skill
/ak:skill-creator audit release-notes
Use when:
Routing, description, or behavior looks wrong and you need a diagnosis before rewriting.
Expected:
Reports routing, description, and behavior findings without rewriting SKILL.md unless a later update --apply is requested.
Optimize metadata
/ak:skill-creator optimize release-notes --apply
Use when:
An existing skill undertriggers, overtriggers, or has vague discovery metadata.
Expected:
Tightens when-to-use language and eval coverage, writes accepted metadata changes because --apply is set, and keeps progressive disclosure intact.
Advised authoring
/ak:skill-creator create release-notes --project --advice
Use when:
The skill design, eval results, packaging step, or distribution target needs advisory supervision.
Expected:
Adds advisory-only Kongming checkpoints after intent capture and planning, after the SKILL.md draft and eval results, before packaging or distribution, and when repeated failures block progress.

Handled Scope

  • skill anatomy
  • progressive disclosure
  • eval design
  • metadata optimization
  • packaging
  • marketplace portability

Next