prompt-writer

Prompt Writer

You are about to write or edit a prompt another Claude will execute. This skill is the playbook: rules, architecture, snippets, and worked examples for producing a high-signal prompt on the first try. Primary target is Claude Opus 5; the same patterns work on Sonnet 5 at lower cost and on Fable 5 above it.

Skim this body, jump to the reference that matches the task, fill in the template, validate against the checklist. The body carries the workflow; the references in ${CLAUDE_SKILL_DIR}/references/ carry the depth.

Ground before you write

Do not author a prompt from your own built-in knowledge alone. Parametric knowledge is frozen at the training cutoff; the world the prompt runs in is not. Two grounding passes come before the first line, every time.

• Read the actual target. When the prompt names an existing file, agent, skill, command, or symbol, open it first with Read, Grep, Glob, or ac:explore. Editing an existing prompt means reading its current body and the references it points at, not recalling what it likely says. A prompt built on a guessed file shape ships the guess.
• Verify against fresh sources. Model behavior, effort and thinking parameter shapes, API surfaces, SDK signatures, and library or framework features shift between versions, so built-in knowledge goes stale (a Laravel Horizon balancing option you "know" from the training cutoff may have changed in the latest release). Confirm anything version-sensitive against current canonical docs through ac:librarian or the ac MCP web tools (web-search, web-fetch, search-docs) before you state it as fact, and cite what you found.

When your built-in knowledge disagrees with the file or a fresh source, the file and the source win.

Decision flow

Route by what you are about to produce. Each branch points at a reference for the depth.

Writing a SKILL.md body (any Claude Code skill)?
├── YES → use this skill for the prompt body, route the skill SHAPE
│         (frontmatter, scope, invocation, bundling) through `skill-creator`.
└── NO  → continue
        ↓
Writing the body of a slash command (/name [args])?
├── YES → use this skill for the prompt body, route the command SHAPE
│         (arguments, allowed-tools, shell injection) through `command-creator`.
└── NO  → continue
        ↓
Writing a custom subagent definition (.claude/agents/<name>.md)?
├── YES → use this skill for the system-prompt body, route the agent SHAPE
│         (tools, model, permissions, isolation) through `agent-creator`.
│         Then read `${CLAUDE_SKILL_DIR}/references/subagent-prompts.md`.
└── NO  → continue
        ↓
Writing CLAUDE.md, CLAUDE.local.md, or .claude/rules/*.md?
├── YES → use this skill for content tone and structure, route the file SHAPE
│         (scope, paths:, @imports, loading order) through `claude-md-rules-creator`.
└── NO  → continue
        ↓
Briefing a fresh Agent tool call (no custom subagent_type)?
├── YES → `${CLAUDE_SKILL_DIR}/references/subagent-prompts.md`
└── NO  → continue
        ↓
Auditing or improving an existing prompt?
├── YES → `${CLAUDE_SKILL_DIR}/references/anti-patterns.md` FIRST,
│         then `${CLAUDE_SKILL_DIR}/references/architecture.md`.
└── NO  → use the Quick template below; validate against the Quick checklist.