Skill Creator
inherit skill authoring bounded scopeQuick reference
| Field | Value |
|---|---|
| Agent | brewcode:skill-creator |
| Triggers | create skill, improve skill, fix activation |
| Model | inherit: uses the session model |
| Tools | Read, Write, Edit, Glob, Grep, Bash, Agent |
| Output | Skill directory, artifact list, validator/evaluation results, and remaining questions |
What it does
Create a reusable Claude Code skill from a concrete workflow, improve an existing skill, or apply confirmed review findings. Skill Creator chooses the structure, writes instructions and supporting resources, validates the result, and evaluates realistic examples.
Start with a clear goal, exact directory ownership, relevant completed work, the next consumer, and acceptance evidence. The agent follows the closest well-built skill in the repository, alongside conventions and project rules. It handles one bounded deliverable; larger requests return a split proposal before implementation.
When to use
- Capture a proven workflow as a reusable skill.
- Add a skill with deterministic scripts and supporting references.
- Fix incorrect frontmatter, tool assumptions, or invocation behavior.
- Apply independently confirmed review findings to an existing skill.
Example
Ask the main session to delegate:
Create one project skill for database migrations. Own .claude/skills/db-migration/.
Follow the existing migration workflow and project test commands. Require explicit
invocation; return artifact paths, validator verdict, and output-evaluation evidence.Expected result: a SKILL.md and only the references, scripts, tests, and README needed by that workflow. Detailed logs stay in a report file; the final return names written artifacts, validation results, and unresolved decisions.
Workflow
- Resolve scope and workflow
Extract a proven workflow from available conversation context or the caller’s brief. Resolve project, personal, or plugin location. Ordinary subagents return unresolved questions to the caller instead of planning an interactive question round.
- Choose invocation and execution context
Distributed Brewcode skills require user-invocable: true and disable-model-invocation: true. Choose forked or inline execution for the actual workflow, and verify any selected custom agent exists.
- Write the prompt contract and resources
Use a free-text prompt entry point, explicit mode detection where needed, and a PLAN block before actions. Keep instructions concise; load schemas and detailed modes from references on demand. Use scripts for deterministic work.
- Validate and evaluate
Run validate-skill.sh, fix confirmed failures, and rerun it after changes. Evaluate three to five realistic prompts against fresh-session baselines. For explicit-only skills, evaluate output through explicit invocation instead of measuring automatic activation.
- Finish with tests, documentation, and caller-owned optimization
Run meaningful script tests when scripts are present. Write a concise README with real examples. Request the final text-optimizer pass from the main caller; the creator does not nest delegation or delegate its own verification.
Invocation policy
| Context | Configuration |
|---|---|
| Distributed Brewcode product skill | user-invocable: true and disable-model-invocation: true |
| User-only skill outside distributed products | disable-model-invocation: true |
| Model-only knowledge outside distributed products | user-invocable: false |
| Both invocation paths outside distributed products | Default invocation behavior |
The distributed-product rule is repository policy, stricter than upstream defaults. Explicit invocation makes the entry point deliberate; it does not guarantee correct execution. Claude Code publishes no automatic-activation percentage.
Evaluate behavior, not invented activation rates
Description and trigger phrasing can affect discovery for model-invocable skills. Use actual triggered/missed and output results from realistic runs. For disable-model-invocation skills, automatic activation and subagent preloading are unavailable by design.
Execution and tools
A skill fork is an ordinary subagent, not a conversation fork. A skill with
context: fork starts an isolated subagent. Ordinary subagents have tool filtering and cannot
use AskUserQuestion for an interactive user round. Conversation forks retain the parent’s tools;
do not apply ordinary-subagent limitations to that distinct mode.
Upstream supports configurable nested-agent depth, but this project’s authoring policy keeps delegation with the main caller. Skill Creator completes bounded work itself and returns split or optimization requests to the caller. Agent being listed as a tool does not authorize ignoring that policy.
allowed-tools preapproves supported calls; it is not a tool-access sandbox. Grants and
disallowed-tools restrictions apply to the invoking turn and clear on the next user message,
while skill instructions persist. Scope permissions to the required calls instead of
preapproving every Bash, Write, Edit, or Agent action.
Dynamic skill-text substitutions such as ${CLAUDE_SKILL_DIR} are not ordinary environment
variables available to hooks or agents. Use the documented substitution and resource path for
the execution surface being authored.
Validation and references
The creator uses the bundled validator for frontmatter, naming, prompt contract, structure, and other supported checks. A passing validator complements evaluation; it does not prove the workflow achieves its goal. Confirmed fixes require another validator run.
Reference files under brewcode/skills/skills/references/ provide the current frontmatter catalog,
execution model, design patterns, activation troubleshooting, prompt contract, and review checklist.
The source was refreshed against upstream contracts through Claude Code 2.1.285; historical line
citations in references are context, not a permanent API guarantee.
The agent checkpoints artifact paths and status in its report directory as it works. Its maxTurns limit is a loop stop; partial work remains incomplete and must be resumed from evidence.
Related
Brewcode overview
Skills
Hook Creator
Agent source
Authoring references
Updating plugins
/brewtools:plugin-update to check and update the brewcode plugin suite in one command.
See the FAQ for details.