Agent Instructions is a skill for writing or restructuring an app repository’s entry file, adding a rule when you say “remember this”, and fixing the file after an agent ignored a rule it had. It treats the entry file as a router: a short file that points to operative files by the task the agent is doing.
It also keeps the file healthy over time. Every rule carries a dated receipt, files have word budgets held by a script, and two passes (at each release and after each mistake) keep it from growing into a manual.
What You Get
- A starter AGENTS.md skeleton with a preamble, a router and a product section
- A router pattern with rows worded the way a user would say the task
- Two scripts: one lists docs no chain from the entry file reaches, one checks word budgets and rules past their review date
- Working rules for the file, each with a receipt in a separate receipts file
- A release pass and a mistake pass for keeping the file current
Common Use Cases
- Creating the first AGENTS.md for a new repository
- Restructuring a bloated CLAUDE.md into a router with area files
- Turning a corrected mistake into a dated rule, or into a test or script when memory would fail
- Checking that Claude Code and Codex both read the same master instruction file
What Makes It Unique
Most guidance on instruction files lists rules to include. This skill is about when a rule reaches the agent: router rows keyed by moment, a cap on always-on rules, a word budget per file, and the test of asking a fresh session to quote a rule only the file holds. A rule that depends on memory at a decisive moment is turned into a script, test or required argument. Every rule keeps the date and the failure that taught it.
Under The Hood
The skill is mostly instructions the model reads, with five reference files and a receipts file loaded on demand. It ships two Python scripts that use only the standard library and read the repository without writing anything: one walks the links from the entry files to find unreachable docs, the other counts words per file against a budgets file. The agent runs them with shell commands and edits the repository’s instruction files and docs. It fetches no web pages and calls no APIs.