Giving AI the Right Context

Even a perfectly worded prompt fails if the AI doesn't know your codebase's conventions, constraints, or history. Context often matters more than phrasing.

What counts as context

• Relevant files, pointing at the actual code involved instead of describing it from memory.
• Project conventions, naming patterns, folder structure, preferred libraries, things a new human hire would also need to learn.
• Error messages and logs, the exact text, not a paraphrase, stack traces often contain the one detail that matters.
• Constraints, "don't add new dependencies," "this must stay backward compatible," "keep this file under 200 lines."

Standing instructions

Many AI coding tools support a persistent instructions file (Claude Code uses CLAUDE.md, for example) that's automatically included in every conversation, a good place for conventions that apply to every task, so you don't have to repeat them.

# CLAUDE.md
- Use Tailwind utility classes, not custom CSS files.
- All new API routes need a matching Zod schema.
- Run `npm test` before considering a change complete.

Breaking work into steps

Large, multi-part tasks ("build a whole checkout flow") produce weaker results than a sequence of smaller, well-defined ones ("add the cart total calculation," then "wire it into the checkout page," then "add validation"). Each smaller step is easier to verify, and a mistake in step one doesn't compound through steps two and three before you notice it.