Prompt Authoring Playbook

---
name: prompt-authoring-playbook
description: Use when WRITING or FIXING any prompt — for an app, an agent's system prompt, an eval judge, or another model — the five-part prompt contract, the build procedure, anti-patterns with fixes, and the meta-prompting repair loop.
---

# Prompt-Authoring Handbook — v2026-07-07 · queue: FN-6

## 0. Scope / when to ignore this
Writing prompts that other code or people will rely on. Ignore for: interactive
one-off chats (just talk), context-window/caching mechanics (FN-5, queued — §6 carries
the two rules that can't wait).

## 1. The prompt contract — every reliable prompt has five parts

| # | Part | What it is | Missing it looks like |
|---|---|---|---|
| 1 | **Goal + success criteria** | What done means, as 3–7 BINARY checks written BEFORE the prompt body | "Improve this text" → improved how, for whom? |
| 2 | **Role + context, at the right altitude** | Enough situation to act; altitude scales with executor (§2 step 4) | A page of backstory the task never uses |
| 3 | **Output contract** | Schema / format / length / language, exactly as consumed downstream | "Give me JSON" with no schema; prose when code parses it |
| 4 | **Delimited input** | User/document data fenced off (tags/fences) and named as DATA | Instructions and data interleaved → injection + confusion |
| 5 | **The uncertainty path** | What to do when the answer isn't determinable: abstain token, ask, or default — CHOSEN, not left to chance | Confident hallucination, because you demanded an answer |

Part 5 is the most-forgotten and highest-value: a prompt that never says what to do
under uncertainty has CHOSEN confident guessing.

**Worked exemplar (all five parts, copy this shape):**