Giving the AI Context

The instinct is to explain your code in prose. Do not. Point at it. A file reference costs you five characters and gives the model the actual truth instead of your summary of it.

Every tool has a syntax for this — @filename in Cursor, referencing paths directly in Claude Code, the attachment picker in Copilot Chat. Learn the one your tool uses on day one.

For most tasks, three things are enough, and they map onto three different questions the model has to answer:

1. 1

   The file you want changed

   Answers “where does this go?”

2. 2

   A similar file that already does it right

   Answers “what should it look like?” A similar-file example gives the model a concrete convention to follow.

   One good example teaches conventions better than a paragraph describing them.

3. 3

   The type or schema it must conform to

   Answers “what shape is the data?” — and stops invented field names dead.

More context is not better. The window is a budget, and filling it with noise pushes out the parts that mattered.

Most tools now accept images, and for UI work a screenshot outperforms any description you could write. “The spacing under the header is too tight on mobile” plus a screenshot is a complete brief.

For errors, paste the whole thing. The instinct to tidy it up is the exact wrong move: the stack frames you trimmed are the part that named the file.

The checkout total is wrong when a coupon is applied. Relevant files:- src/lib/pricing.ts        (where the total is computed)- src/lib/pricing.test.ts   (existing tests, all passing)- src/types/cart.ts         (the shapes involved) Steps to reproduce: add two items, apply SAVE10, total showsthe discount twice. Full error from the console:[paste it all, unedited]

Everything so far assumes the window has room for what you deliberately attach. It has to share that room with something you are not managing directly: the conversation itself.

Every prior message, every file the model already read, every failed attempt sits in the window whether you meant to keep it there or not. A session that has gone through three wrong approaches to the same bug is carrying all three, and by the fourth attempt the model is reasoning with a full history of dead ends crowding out the actual fix.

Before any substantial request, one line of prompt saves entire rounds of confusion:

Do you have everything you need to do this? If anything ismissing or ambiguous, ask before starting.

Models are surprisingly willing to say what they are missing when you invite it. They just never volunteer it, because the default behaviour is to answer.