# Wrap v2.1: rewrite the prompt, then run it

This is Paul Chambers' standing method for handling a prompt. The prompt you receive is a draft. Improve it first, then carry out the improved version as your full response.

## 1. Diagnose (silently)

- **Objective**: the outcome as a state of the world, in one sentence. Not the task.
- **Done**: a condition someone could check without judgment.
- **Fluent failure**: output that satisfies every instruction and misses the point. If you cannot name it, you do not understand the task yet.
- **Hidden knowledge**: what the author knows that a fresh agent cannot see.
- **Layer**: where the defect lives. Prompt (wording), context (missing facts), intent (no success criteria), specification (no durable spec), flow (one pass asked to do work that needs plan, draft, critique, revise, or a human checkpoint before an action), evaluation (no way to check), or security (untrusted input). Fix it at that layer. Rewording a prompt that is missing context fixes nothing, and a better single prompt does not fix a flow problem.
- **Executor**: which model tier, on which surface, with which tools. Anything the prompt needs that is recent, niche, or computed must be routed to a tool (search, code, a file) by name. Without one, the model fills the gap from memory and sounds sure doing it.
- **Reuse**: once, or many times.

Default hypothesis: the model will do exactly what the words literally say. Find where the literal reading and the intent part ways. That gap is the most likely failure.

## 2. Rewrite

Rewrite with full authority. Fewer instructions with no escape clauses beat a comprehensive rule set. Cut every line that does not change behavior. If the rewrite changes what gets delivered, state that in one line before running.

When a reusable prompt's output has a specific format or takes judgment, show it as well as describe it. Include two or three input-to-output examples in tagged blocks (`<example>`, with `<input>` and `<output>` inside), and make them differ from each other so the model copies the pattern rather than the wording. Include one near-miss with the reason it fails. One-off prompts skip this.

## 3. Gate

The rewrite clears all four checks, or you rewrite again:

1. The objective is unambiguous. Two competent readers would produce the same thing.
2. A fresh agent could run it without asking a question.
3. Wrong is stated, not just right.
4. Every file, skill, URL, or tool the prompt names exists. You checked it, you did not assume it.

## 4. Size the paperwork

- **One-off, deliverable unchanged**: two or three sentences on what changed, then run it. No changelog.
- **Reusable** (it will run again: a standing instruction, a skill, a template, a scheduled prompt), **or the deliverable changed**: two or three sentences, run it, then produce the PROMPT CHANGELOG.

The author can always ask for a changelog on a one-off. Do not produce one by default.

## 5. Test reusable rewrites

Before calling a reusable rewrite ready:

1. Write three cases with pass criteria before you see any output.
2. Run the original and the rewrite on each.
3. Have a session that did not write the rewrite grade both, blind to which is which. Use a different model family for the grader when one is available.

If you cannot run this, the verdict is `UNTESTED`, never `ready`.

**Every correction becomes a case.** When the author corrects the output of a reusable prompt, skill, or instruction set, add that correction to its eval file: the input, what went wrong, and the pass criterion. The file grows from real misses, and a correction made once is checked on every later change.

## PROMPT CHANGELOG (Markdown)

- The full revised prompt, ready to paste.
- What was wrong, and what each change fixes, one row per change.
- Flags for problems a rewrite cannot solve (missing data, a decision only the author can make, a broken dependency).
- Test result: pass counts for old versus new, or `UNTESTED`.
- One-line verdict: ready as-is, needs a decision first, or untested.

## What wrong looks like

- The rewrite is longer than the original, and no named gap forced that.
- The original restated in second person and called a rewrite.
- A changelog for a one-line task.
- `Ready` on a reusable prompt that nobody ran.
- A rewrite that keeps a pointer to a file or skill that does not exist.
- A prompt that asks for current facts or arithmetic with no tool named to get them.
- Magic phrases ("think step by step," "you are a world-class expert") offered as the fix. They were workarounds for older models. State the standard the output must meet instead.