Builder

---
name: builder
description: |
  Author and refine reference skills, task-style command skills, and path-scoped
  rules using one repository-grounded workflow. TRIGGER when: asked to create,
  build, scaffold, convert, review, or update a skill, command, workflow, rule,
  coding standard, or contextual instruction. Skills are the canonical primitive
  for a reusable role, procedure, or body of judgment — there is no project-agent
  artifact type.
argument-hint: "skill|command|rule <name-or-request>"
allowed-tools: Read, Write, Edit, Glob, Grep, Bash
---

# Builder

Build one artifact through the matching type reference. Run inline and inherit the
session model; do not fork or override the model unless the artifact being authored
has an independently justified need.

Arguments received: `$ARGUMENTS`

## Dispatch

1. Treat the first whitespace-delimited argument as `TYPE` and the remainder as
   the artifact name, request, or path.
2. Accept exactly these types:

   | Type | Read and follow |
   |------|-----------------|
   | `skill` | `references/skill.md` |
   | `command` | `references/command.md` |
   | `rule` | `references/rule.md` |

3. If `TYPE` is missing or unknown, or the remaining request is empty or only
   whitespace, print the following and stop without reading type references or
   modifying files:

   ```text
   Usage: /builder <skill|command|rule> <name-or-request>
   ```

   `agent` is not an artifact type. A reusable role, procedure, or specialist
   judgment is authored as a skill; a bounded isolated worker context is an
   execution choice made by `/delegate`, not a repository artifact.

4. Read the selected reference completely, then execute its protocol against the
   remainder of `$ARGUMENTS`. The selected reference is authoritative for artifact
   shape and type-specific validation.

## Shared protocol

Apply these steps for every valid type before the selected reference's type-specific
steps.

### 1. Discover local authority

- Find and read applicable `AGENTS.md` and `CLAUDE.md` files from repository root
  through the target directory. More local instructions win; in one directory,
  `AGENTS.md` is canonical.
- Identify the source-of-truth artifact directory. In Open Harness and equipped
  projects, edit `.oh/skills/`; provider directories such as `.claude/`,
  `.codex/`, and `.pi/` are generated or symlinked exposure surfaces.
- Outside an Open Harness layout, follow the target project's documented canonical
  path rather than creating `.oh/` speculatively.
- Inspect two or three nearby artifacts of the same type. Reuse their naming,
  frontmatter, structure, tone, and validation conventions.
- Search for an existing artifact with the same purpose. Prefer a focused update or
  explicit consolidation over a near-duplicate.
- Consult compiled harness patterns before proposing a change. Run
  `/wiki query <artifact-name-or-subsystem> --patterns` and read what it returns:
  each page records a failure mode, its root cause, and a workaround this harness
  already paid for. Cite the motivating `[[pattern-...]]` slugs in the report.
- Read `.oh/evals/decisions/skill-impact.md` for prior proposals against the same
  target. Do not re-propose a change recorded there as `REJECTED` unless new evidence
  contradicts the recorded validation; when you do, cite the prior record id.

### 2. Define the contract

Before editing, state internally:

- the artifact's one-sentence purpose and concrete triggers;
- who invokes or consumes it;
- what is in scope, out of scope, and considered done;
- the minimum tools, context, side effects, and supporting resources required;
- which behavior is repository-specific and must be grounded in inspected files.

Ask a question only when unresolved ambiguity would materially change the artifact
or create unsafe side effects. Otherwise use the request and repository evidence.

### 3. Author narrowly

- Use lowercase kebab-case names and one artifact per coherent concern.
- Put matching and trigger information in frontmatter, not only in the body.
- Use imperative, operational language. Remove generic expertise prose that does
  not change behavior.
- Prefer the least privilege and smallest context footprint that completes the job.
- Cite real local paths and commands only after verifying them.
- Do not modify unrelated files, generated provider mirrors, or user work in the
  working tree.

### 4. Validate and report

- Validate frontmatter delimiters and required fields without assuming optional YAML
  libraries are installed.
- Check every referenced path, invocation, tool, and supporting file.
- Enforce the selected reference's size, safety, and semantic checks.
- When `.oh/scripts/link-providers.sh` exists and canonical `.oh/` primitives were
  changed, run `bash .oh/scripts/link-providers.sh --check`.
- Run `git diff --check` when inside a Git worktree.
- Report the files created, updated, or removed; the resulting invocation or loading
  behavior; key design choices; and validation evidence. Never claim a check ran if
  it did not.
- Append a `PROPOSED` record to `.oh/evals/decisions/skill-impact.md` when a skill
  edit lands: the next `SI-nnnn` id, the one-sentence proposal, the single
  target artifact, the motivating pattern slugs, and the unified diff scoped to that
  target path. `motivating patterns: none (direct request)` is a legitimate value —
  record it rather than inventing a pattern to cite. Stage the ledger
  (`.oh/evals/decisions/` is tracked, so a plain `git add`) and report the
  allocated id. Never edit an existing record.