# The Outcome-Description Prompt (Chapter 9)

**Use when:** you are pointing your agent at the report factory for the first time, or re-grounding it after drift.

Describe the outcome in your own words, Chapter 5 style: what the report is for, what must be true when it lands, and what you will check before you ship. Attach your style file if you have one; the style-file hook is what keeps every draft sounding like you instead of like a template.

```
Here is the report's outcome, in my own words:
[OUTCOME: WHAT THE REPORT MUST MAKE TRUE FOR ITS READER. "The director can
report the state upward without calling me." Not "a monthly report."]

The chain:
Pull: [INPUT FOLDER]
Analyze: [WHAT CHANGES GET COMPUTED, IN PLAIN WORDS]
Artifact: [FORMAT, HOME, WHO READS IT]

My style file is attached. Write in the voice of the style file: same
facts, same ask.

Draft only. Nothing ships without my review. Before you draft, explain
the delta logic back to me in two sentences, so I can check you
understand what to compute.
```

## The style-file hook

If you keep a style file (the block lives in templates/AGENTS-blank.md), one line carries it into every factory run: "Write this in the voice of the style file. Same facts, same ask." Voice is a checklist, not a feeling, and the style file is the checklist.

## The first-run recipe

1. Send the outcome description with fake data that matches the shape. Real data never enters the prompt; the shapes travel, the data stays.
2. Run on demand until two drafts pass your review.
3. Make it explain the delta logic back to you, then run once more when it is next due.
4. If it is good, schedule it. If it is not, you have learned exactly what is worth automating.
5. The input folder is the only input, and the first three runs happen while you watch.

Verify: every output arrives as a draft, the style file shows in the voice, and the review gate is you.