Writing Commands

Announce: "Using writing-commands skill for command creation, editing, or review."

Invariant Principles

1. Commands are direct prompts: A command loads entirely into context. No phases dispatch to subagents. No orchestration layer. The agent reads it and does the work.
2. Structure enables scanning: Agents under pressure skim. Sections, tables, and code blocks catch the eye. Prose paragraphs get skipped.
3. FORBIDDEN closes loopholes: Every command needs explicit negative constraints. Agents rationalize under pressure. Each excuse needs a counter.
4. Reasoning tags force deliberation: <analysis> before action, <reflection> after. Without these, agents skip straight to output.
5. Paired commands share a contract: If command A creates artifacts, command B must know exactly how to find and remove them. The manifest/contract is the interface.

Inputs

Outputs

Phase Overview

Phase 1: Create Command

Define command structure using the schema: file naming, frontmatter, required sections, optional sections, and token efficiency targets.

Execute: /writing-commands-create

Outputs: Command file at commands/<name>.md

Self-Check: Frontmatter present, all required sections included, imperative language used, token targets met.

Phase 2: Review Command

Run the quality checklist against the command. Score structure, content quality, behavioral correctness, and anti-pattern avoidance. Follow the review and testing protocols.

Execute: /writing-commands-review

Outputs: Review report with score, passing/failing checks, critical issues.

Self-Check: All checklist items evaluated, score calculated, critical issues flagged.

Phase 3: Paired Commands

When a command creates artifacts, ensure a paired removal command exists with proper manifest, discovery, safety, and verification contracts.

Execute: /writing-commands-paired

Outputs: Paired command file, cross-references in both commands.

Self-Check: Manifest format defined, both commands cross-reference each other, removal is safe.

Self-Check

Before completing command creation or review:

• Frontmatter has description field with triggers, not workflow
• MISSION section is one clear paragraph
• ROLE tag has domain expert + stakes
• 3-5 Invariant Principles, each testable
• Execution steps are numbered and imperative
• Every step that can fail has a failure path
• Output section has concrete format
• FORBIDDEN section has 5+ specific prohibitions
• Analysis tag prompts pre-action reasoning
• Reflection tag asks specific verification questions
• If paired: partner command referenced, manifest format defined

If ANY unchecked: STOP and fix before declaring complete.

<FINAL_EMPHASIS> Commands are the atomic unit of agent behavior. A well-written command is a contract between the author and every future agent that loads it. Ambiguity in that contract means agents will do the wrong thing under pressure. Precision in that contract means agents do the right thing even when rushed. Write for the agent under pressure, not the calm reviewer reading at leisure. </FINAL_EMPHASIS>