Context in a Minute: Project instructions

In one sentence

Project instructions are checked-in rules that tell coding agents how this repository expects work to be done.

Why it matters

Chat memory is fragile. Repos are durable. If your test command only lives in someone’s head, agents will invent a wrong one.

How it works

Put operational truth in instruction files (CLAUDE.md, AGENTS.md, or your tool’s equivalent):

• commands to install, run, test, lint;
• hard constraints;
• pointers to deeper docs.

Keep narrative design docs separate. Agents need the checklist more than the essay.

Review instruction files in PRs like code. They change agent behavior for everyone.

Example

Good instruction line:

Before claiming done, run npm test and npm run lint.

Weak instruction line:

Be careful and write clean code.

Agentic coding use

Project instructions are the base layer of context engineering. Task prompts add the temporary goal; instruction files supply the permanent physics of the repo.

Watch out

Duplicate rules across tools will drift. Prefer one canonical section and link from tool-specific files when needed.