write-prompt

Write Prompt

Covers what Claude skips without this skill. Everything else about prompt writing, Claude already does: do not pad a prompt with it.

Prompts run in production (LLM API, templates)

A prompt that processes inputs nobody reviews one by one needs all three:

• A worked example. At least one input with its exact expected output. Pick it from the hard cases (messy, multilingual, several items), never a clean one: "Show, Don't Tell: Examples are more effective than descriptions" [1], and an example that does not match the target task misleads ("Example pollution" [1]).
• The exact output schema. Field names, types, formats (ISO 8601, ISO 4217...), allowed values. Say what to output when a field cannot be found (null, never a guess).
• A rule for each edge case the input can produce. None, several, ambiguous, malformed: state the output for each, such as an empty list, one entry per item, or a review flag. "No error handling: Assuming outputs will always be well-formed" is a named pitfall [1].

Rules an agent reads (CLAUDE.md, hooks, skills, commands)

Claude scans bullet and tree structure more reliably than it parses paragraphs. Default to lists over prose for rules, steps, or conditions.

• One item per condition, each stating the trigger and the action, or what is blocked and what to do instead.
• Don't restate what Claude already knows. No explanation of what git, a branch, JSON or a tool Claude already understands is (e.g. what AskUserQuestion does): the reader is a model that knows them.

Persuasion principles

Writing a rule that must hold under pressure (a skill, a hook message, a sub-agent prompt) -> read references/persuasion-principles.md: seven principles (authority, commitment, scarcity, social proof, unity, reciprocity, liking), which to combine per skill type, and their ethical limits.

---

[1] wshobson/agents, prompt-engineering-patterns (MIT). Sources: SOURCES.md.