Command Writing

command-writing

When to use

• Creating a new slash command in src/agent-src/commands/{name}.md
• Rewriting an existing command (not a typo fix)
• Deciding whether a request should be a command at all
• Splitting an oversized command into smaller ones

Do NOT use this skill when:

• The content is a constraint the agent must always honor → use rule-writing
• The content is reference knowledge agents cite → use guideline-writing
• The content is a triggered workflow invoked by the model → use skill-writing

Command vs skill — critical test

A command is user-invoked and carries disable-model-invocation: true. A skill is model-invoked via description routing. If both audiences apply, author as a skill and add a thin command that delegates to it.

A command must EARN a top-level slot. Only three justifications qualify: flow-entry, state-query, or product-surface — everything else is a skill. Check before authoring: command-clusters § Command justification (ADR-048).

Categorize at creation (checklist). When you author a NEW top-level command, declare its justification as a category: in the frontmatter — the schema enum is flow-entry | state-query | product-surface:

• flow-entry — a daily work starting point the user TYPES to begin (work, git-commit, review-changes).
• state-query — a read-only check (agent-status, project-health, profile).
• product-surface — a feature started deliberately (council, research, roadmap).
• fits none → it is a skill. Don't add the command; author a skill.
• genuinely ambiguous? Omit category: and note why — category: is OPTIONAL (validate-when-present); deferring is intentional, not forgotten.

Sub-commands (council:debate) inherit the parent cluster's category — leave category: off them. Why optional + not yet a blocking lint, and what triggers the upgrade: command-category-governance.

Commands ARE Claude skills (projection reality)

Every command in src/agent-src/commands/{name}.md is projected into .claude/skills/{slug}/SKILL.md by scripts/condense.ts (generate_claude_commands). Nested commands flatten with - (council/default.md → council-default). Skills and commands share the same .claude/skills/ namespace — Claude does not distinguish them.

Consequences for authoring:

• The frontmatter description is the routing surface Claude reads. Polite or generic phrasing causes undertriggering even with disable-model-invocation: true set, because the in-host command suggester, fuzzy search, and any tooling that scans .claude/skills/ rank by description match.
• disable-model-invocation: true blocks automatic invocation. It does NOT remove the command from discovery surfaces. A weak description means the command is invisible to the suggester even when the user's intent matches.
• Trigger phrasing must follow the same Iron Law as skill descriptions: name 2+ trigger classes (domains, symptoms, user phrasing), end with the ... even when the user just says ... tail, ≤ 200 chars. See skill-writing § 1b for the canonical before/after.
• The suggestion.trigger_description and suggestion.trigger_context blocks are separate from the frontmatter description — they drive the in-host suggester (command-suggestion-policy), not Claude's skill router. Both matter, both must be precise.

Bottom line: write the command's description as if a skill router will read it — because one will.

Procedure

0. Run the Drafting Protocol

Creating or materially rewriting a command must go through Understand → Research → Draft from the artifact-drafting-protocol rule.

• Understand — what user-facing problem does /{name} solve in one session? What are the inputs, outputs, side effects?
• Research — inspect src/agent-src/templates/command.md, grep commands/ for overlap, and analyze 1–2 peer commands (e.g. create-pr, commit).
• Draft — propose frontmatter (name, description) first, then the step skeleton. Only fill bodies after both are confirmed.

1. Use the template

Canonical source: src/agent-src/templates/command.md.

Minimum frontmatter:

---
name: {command-name}          # must match filename without .md
description: "Short human-readable summary of what /{name} does"
disable-model-invocation: true
skills: [optional-skill-1]    # optional — skills this command delegates to
suggestion:                   # required (road-to-context-aware-command-suggestion Phase 2)
  eligible: true              # default; set false to opt out of auto-surfacing
  trigger_description: "natural-language pattern, comma-separated examples"
  trigger_context: "concrete signal — branch name, file pattern, recent tool output"
---

Or, when opting out:

suggestion:
  eligible: false
  rationale: "one-line reason this command must be invoked deliberately"

Suggestion-block rules (linter-enforced):

• eligible is required and must be true or false.
• eligible: true → both trigger_description and trigger_context must be non-empty (≥ 10 chars each); the linter rejects empty or overly generic patterns. The suggestion layer never auto-executes; the user always picks.
• eligible: false → rationale must be non-empty. Use the opt-out for intentional-only invocations (settings mutations, destructive actions, package-internal tools, niche maintenance).
• Optional confidence_floor (0.0–1.0) and cooldown (e.g. 10m) override the global settings per command.

Eligibility decisions are tracked in agents/settings/contexts/command-suggestion-eligibility.md. Add or revise entries there before changing a command's suggestion block.

When iterating on the description, delegate to the description-assist skill — approval-gated, no silent edits, max two rounds.

2. Structure the body

Required sections in this order:

1. # /{name} heading + one-line summary
2. Source of truth note — works on src/, never on generated directories
3. ## Steps — numbered sub-headings ### 1., ### 2., ...
4. Final step presents findings and asks the user before destructive changes (numbered options per user-interaction rule)
5. Optional ## Rules — short, command-specific constraints

3. Enforce safety gates

• No auto-apply of destructive actions without user confirmation.
• Every step with side effects (git push, file delete, PR merge) asks first.
• If the command calls external APIs, list required keys / permissions.
• If the command edits agent files, target src/ only.

3b. Path conventions in command body

Body links to guidelines / contracts use the verbatim relative form (../../docs/guidelines/..., ../../docs/contracts/...); the condense-time rewriter handles depth. Do not pre-rewrite in source. Do not write src/ in any markdown link target — the file ships into .augment/commands/ and the prefix breaks consumer resolution. The only legitimate src/ strings in a command file are prose mentions and step instructions about where to edit (per § 2 above). Canonical reference: rule-writing § 3b.

4. Enforce the size budget

Normative source: size-enforcement + docs/guidelines/agent-infra/size-and-scope.md.

Commands orchestrate, they do not implement detail. If a step needs a multi-paragraph explanation, extract it into a skill and call it.

5. Validate

• Run ./scripts-run src/scripts/skill_linter src/agent-src/commands/{name}.md → 0 FAIL.
• Run bash scripts/condense.sh --sync → regenerates dist/agent-src/commands/{name}.md.
• Run ./scripts-run src/scripts/condense --generate-tools → creates the Claude symlink at .claude/skills/{name}/SKILL.md.
• Run the full CI pipeline locally (see Taskfile.yml in this repo for the script list) — must exit 0 except for tolerated warnings.

6. Governance baseline (when introducing a new linter check)

Advisory, reviewer-checked — no CI gate. When the same PR adds a new check to scripts/skill_linter.ts (or strengthens an existing one) such that previously-clean commands now warn, the PR body MUST record the pre-existing violations on main in a Markdown table:

### Pre-existing baseline (informational)

| Code | Count on main | Bucket |
|---|---:|---|
| {new_code} | N | (a) genuine fix · (b) accept · (c) check too aggressive |

Forward-only: the new check applies to the file under review and to future edits. The baseline table is informational so reviewers can spot intent (fix-now vs. backlog) without diffing the full lint output. See agents/evidence/analysis/lint-warning-triage.md for the 3-bucket reference.

Output format

1. Complete command file at src/agent-src/commands/{name}.md
2. Frontmatter populated, disable-model-invocation: true present
3. Linter output showing 0 FAIL
4. Generated Claude symlink verified

Gotchas

• Forgetting disable-model-invocation: true — the model will auto-invoke the command as if it were a skill.
• Numbered options without a "skip" / "no change" path.
• Steps that silently apply destructive changes — always show summary + ask.
• Referring to .augment/ paths for editing — source of truth is src/.
• Duplicating another command's workflow instead of delegating via skills:.

Frugality Standards

Apply the Frugality Charter to every command you author.

Examples in this artifact:

• Per the charter's default-terse rule, command output blocks state the action result, not "Now we will execute…".
• Per the post-action summary suppression, the success path emits the artifact (PR URL, commit hash) without a wrapping summary.
• Per the cheap-question check, never offer "preview vs. execute" as a numbered option when the command's role is to execute.

Pre-save self-check:

1. Does any command step prescribe a "Let me…" or "Found it" output line?
2. Does the command default to multi-line summaries when a one-line outcome suffices?
3. Is a confirmation gate used outside the Iron-Law / Routine / Contextual taxonomy?
4. Are template placeholders ({{var}}) accompanied by setup prose instead of action prose?

Do NOT

• Do NOT set disable-model-invocation: false
• Do NOT auto-apply destructive actions
• Do NOT inline skill-level detail — delegate
• Do NOT edit dist/agent-src/, .augment/, or .claude/ projections
• Do NOT exceed the hard size limit without a waiver

Cloud Behavior

On cloud surfaces (Claude.ai Web, Skills API) the package's scripts/skill_linter.ts, scripts/condense.ts, and the task runner are not available. This skill still applies — but with prose-only validation:

• Emit the full command file as a copyable Markdown block. Do not attempt to write it to disk.
• Self-check the frontmatter against the rules below — name, description, disable-model-invocation: true MUST all be present.
• Self-check the body shape: numbered steps, explicit safety gates, no inline skill-level detail.
• Tell the user to save the file under src/agent-src/commands/{name}.md and run task sync && task lint-skills locally before committing.
• Skip every reference to running the linter, condenseor, or generators yourself — they only run on the user's machine.

Examples

Good description (trigger-shaped, outcome-focused):

"Create a GitHub PR with structured description from Jira ticket and code changes"

Bad description (vague, no outcome):

"PR command"

The ## Examples section on a user-facing command

Stated once, here — never repeated per command.

A command whose visibility is visible or advanced carries an ## Examples section: 1-3 filled invocations in a fence, then exactly one Why line. Filled means a real path, a real ticket key, a real flag — a placeholder teaches nothing about what to type.

The Why line uses one of two literals, **Why it works:** or a ### Why it works heading, and cites exactly one pattern id from prompt-patterns.yml: outcome-not-steps · self-check-loop · measurable-target · give-the-artifact · point-at-reference · say-the-format. One id, because the Why line's job is the part that transfers to a command the reader has not seen, and a line naming three patterns names none of them.

## Examples

```
/work fix the failing login test under tests/feature/auth
```

**Why it works:** measurable-target — it names the file instead of the symptom.

Enforced by check_command_examples, forward-only: the 23 in-scope commands as of 2026-08-24 are grandfathered in a frozen set inside the gate, and a command added or promoted into visible/advanced after that must comply. Two things the gate checks that are easy to miss — an invocation must resolve to the command's own name: (a rename otherwise leaves stale examples behind), and a flag used in an example must be documented in the body outside the section.

Encode usage policy in the description

Workflow sequencing, preconditions, ID/output provenance ("copy ids verbatim, never from memory"), a mandatory "why" intent field, and turn-end contracts belong INSIDE this artifact's description/frontmatter — where they fire at the decision point — not in always-on prose. See tool-description-as-policy.