craft-prompt
craft-prompt
Purpose
Turn a goal, scattered notes, or a raw ask into a prompt the user copies as-is into another model, agent, or product surface. The prompt is the product: when the user asked for one, deliver it even if the underlying task looks simple enough to do directly.
Workflow
Infer what the request and available context already establish; ask only when the answer would materially change the prompt.
- Outcome. State what should be true or delivered at the end. Keep a shape, length, tone, or language only when the user named it or the consumer requires it.
- Context. The prompt usually travels without this conversation, so include the facts the target cannot infer or retrieve; omit what its surface already provides through files, tools, or attachments. In coding or worktree prompts, paths are worktree-relative unless the user needs machine-specific ones.
- Boundaries. Add scope, non-goals, approval limits, or irreversible-action rules only where crossing them would matter, stated once.
- Evidence. Where done could be read more than one way, say what observable check shows it is done. Don't add generic "be thorough" caution.
- Controls on demand. A role, format contract, example, XML boundary, process step, or tool rule goes in only when the consumer requires it or it corrects a likely or observed failure —
references/components-guide.mdis the menu. Use the target's native controls (effort, verbosity, structured output) rather than simulating them in prose.
Special cases: for a /goal, read references/goal-conditions.md first. For a reusable template, turn each value that genuinely varies into a {{placeholder}}; bake specifics into a one-shot. For image, video, or system prompts, start from the matching file in templates/.
A one-sentence prompt for a simple task is a feature. If removing a line would not change a competent target's output, cut it.