Two ways to use this
- As a reference — the rules for writing any prompt, CLAUDE.md, or post an LLM reads.
- As a drop-in — paste the block at the bottom into a prompt, a system prompt, or your CLAUDE.md, and the model formats to these rules.
The one law
Models do not read, they parse. They scan for boundaries and weigh position.
Same model, same document: moving one line from the middle to the edge took recall from 27% to 98%. Everything below follows from that.
1. Draw boundaries the model can find
- Separate the kinds of information: instructions, context, examples, each in its own block, never fused into one paragraph.
- Light walls: Markdown headings (
# Instructions,# Context,# Examples). - Hard walls or mixed content: XML tags (
<instructions> … </instructions>). - Pick one structure and stay consistent across the input.
2. Put what matters where the model looks
- Most important instruction first.
- Context first, your question last.
- Long input? Repeat the key ask at the end. Edges get read; the middle gets skimmed.
- Never strand a critical line in the middle. Lift it to the top or wrap it as its own block.
3. Pick the format for the job
| Format | Use it for | Watch out |
|---|---|---|
| Markdown | prompts, CLAUDE.md (talking TO the model) | lightest; nest XML when you need harder walls |
| XML | wrapping mixed context + examples | skip if the wrapped text is already full of XML |
| JSON | structured data in or out (exchanging WITH the model) | verbose; not a container for many documents |
| YAML | config-style data | easy to read; keep it simple |
| HTML | publishing so AI can find and cite you | strip to Markdown when you FEED a page to AI |
Rule of thumb: Markdown / XML to instruct · JSON / YAML to exchange data · HTML to be discovered.
4. CLAUDE.md and agent files
- Most important rule on line one.
- Keep it lean. Per-rule adherence drops as rules pile up (about 68% at 500 instructions); aim under ~200 lines.
5. Word it as do, not don't
- “Use sentence case” beats “do not use title case.” A “do not X” still puts X in front of the model.
6. To get cited by AI (when you publish)
- Answer first, in self-contained 40-75 word chunks.
- Use tables (cited roughly 2.3x more often), how-to and FAQ structure, comparison lists.
- Plain question headings that state what the section answers.
- Publish clean semantic HTML with structured data underneath.
7. Don't strangle the output
- Let the model reason first, then format. A rigid output schema imposed too early can cost it reasoning room.
Paste this into your AI
Format everything you write, and read everything I give you, by these rules. Parse for structure and position, not prose. Separate instructions, context, and examples into labeled blocks (Markdown headings, or XML tags for mixed content), and stay consistent. Put the most important instruction first; in a long input repeat the key ask at the end; never bury a critical line in the middle. Use Markdown or XML to instruct, JSON or YAML to exchange data. In a CLAUDE.md, most important rule first, keep it lean. Phrase rules as what to do, not what to avoid. When writing for publication, answer first in short self-contained chunks, use tables and plain question headings.
Jenny Ouyang · buildtolaunch.substack.com · from “LLMs Don't Read, They Parse”