Give the Right Context

Last week you and the assistant agreed that Kitty stores money as whole paise. Today, in a new conversation, it writes a function that uses floats. It is not being stubborn. That decision simply is not in front of it. An assistant works only from what is in its context: your prompt, the files it has been shown or has read, and the earlier messages in the current session.

The rule

Give the assistant what this task needs, and nothing it doesn't.

The assistant does not know your flatmates, your decisions from last week or the error on your screen unless they reach its context. Useful context usually comes in four kinds:

• The relevant files. The function in question, its tests, and the table layout it reads. Not the whole project.
• Exact evidence. The full error message or the exact wrong output, copied, not "it's broken".
• Rules and decisions. "Money is whole paise." "No JavaScript framework." "Do not change the tests."
• What you have tried. So it does not suggest the same thing again.

More is not always better. Unrelated files can distract the assistant and bury what matters. Old versions of a file, still sitting earlier in a long conversation, can mislead it. And some things must never go into context at all: real passwords, API keys and personal data. Share the error, not the secret.

The shortcut

Before you send, imagine a skilled stranger who has never seen Kitty. Ask: what would they need to see to do this task? List the files, evidence, rules and attempts, then cross off anything they would not need.

That last line narrows the search from "anywhere" to "the query", which is exactly what a skilled stranger would want to know.

Drills

Answer fast — trust the rule.

Remember

• Rule: give the assistant what this task needs, and nothing it doesn't.
• Shortcut: picture a skilled stranger. Files, evidence, rules, attempts, then cross off the rest.
• Trap: assuming the assistant remembers decisions that are no longer in its context.