writing-for-agentslisted

# Writing for agents — the document is a process, not a text - **Can:** restructure, split, prune and rename any document in this kit. - **Must:** state which of the two loads each addition spends, and keep one meaning in one place. - **Cannot:** add a rule without a reason attached, or leave a prohibition standing without the positive behaviour beside it. A document for a human is judged by what the reader understands. A document for an agent is judged by **what the agent does on a run it has not had yet**. Those come apart: a paragraph that reads well and changes no behaviour is a cost with no return, and it is the most common thing in a bad skill file. ## The two budgets, and which one each addition spends | | What it costs | Who pays | |---|---|---| | **Context** | tokens and attention on **every turn**, whether or not the material fires | the agent | | **Attention of the human** | remembering that the document exists and when to reach for it | you | A skill's `description`, a line in `AGENTS.md`, anything always loaded: context. Material behind a pointer costs only the pointer's own line. Material with no pointer at all costs nothing in context and everything in memory — the human is the index. 🔴 **Neither budget is to be minimised.** Human attention is the price of human agency: spend it where judgment matters, and take it back where it does not. ### Invocation is the lever on the first budget In Claude Code, two frontmatter fields decide who can reach a skil