Writing for other agents
Content that agents will consume (issue bodies, specs, CLAUDE.md entries) benefits from:
- •Grounded over speculative — state what happened or exists, not what might be useful
- •Knowledge over instructions — inform, don't command. Agents have their own task instructions
- •Skimmable over comprehensive — clear headers for different reader tasks
Check clarity with subagents
When you produce content other agents will consume (CLAUDE.md files, mathematical definitions, proofs, algorithm descriptions): test comprehension by having a fresh Sonnet subagent attempt to USE the content (e.g. "implement from this description" or "answer these specific questions about the algorithm") as a thought experiment (e.g. "answer with your plan but make no edits"). Check whether their output matches your intent. Do not ask "is this clear?" — an agent that misunderstands may confidently say yes.
Editing CLAUDE.md
Agents edit CLAUDE.md directly on their branch. Jörn reviews via git diff in VS Code and merges.
When Jörn edits a file mid-session, immediately commit or snapshot his version before making further changes. This gives you a clean diff baseline to distinguish your changes from his.
The file follows structural rules to prevent information-destroying edits. Read CLAUDE.adr.md for the full style guide and the reasoning behind each rule. Key principles:
- •One claim per bullet. Dense prose packs multiple claims that get lost when a sentence is rewritten.
- •Qualifier preservation. Every adjective narrows meaning. "Clear, detailed, explicit, structured" is not a synonym list — each word names a different quality bar. When rewriting, check: does this rewrite preserve all constraints the original imposed?
- •Clarity & unambiguousness > correctness > maintainability >>> tokens. Redundancy is welcome. Using 50 extra words to prevent a misunderstanding is always worth it.
Content conventions:
- •Invariants and behaviors are documented only after empirically confirmed as useful
- •Label invariants as
[aspirational]if not yet satisfied - •Prefer simple, common, expected rules that don't claim excessive agent attention beyond their assigned work