| name | directing-your-ai-assistant |
| description | Teaches the working method for using an AI coding assistant on analysis tasks — one task per fresh conversation, search before assuming, interview before big plans, validate everything, commit per finished task. Use at the start of any assignment, whenever a conversation gets long and confused, or when the assistant starts guessing. |
Directing your AI assistant
You are the engineer. The assistant is a fast, tireless drafter. This skill is how you stay in
charge — because an assistant with a confused human produces confused work, fast.
The six rules
1. One task, one fresh conversation. An assistant's memory (its "context window") fills up
and its quality drops as it fills. So: finish one assignment section per conversation, then start
a new one. Never drag one giant chat across a whole week — old confusion bleeds into new work.
2. Load only what the task needs. Point it at the router (skills/README.md), load the one
skill for this task, and stop. More instructions ≠ better output — past a point, more is worse.
3. Search before assuming. Before the assistant writes anything, make it look: "Search this
repo first — don't assume something is missing or not implemented." Assistants invent functions
and columns when they haven't looked. Anchoring them in real files kills most hallucinations.
4. Interview before big plans. For anything bigger than one cell, say: "Before you write
code, interview me — ask the questions you need answered." Wrong assumptions caught in questions
cost seconds; caught in finished code they cost hours. Do not outsource the thinking.
5. You validate — always. The assistant drafts; you check. Every task ends the same way:
run the notebook top to bottom (Runtime → Run all), read the output, and ask "does this actually
answer my question?" An unverified answer is not an answer. Checks are your safety rail: the more
things that can automatically reject bad output (a rerun, an assertion, CI), the safer you move.
6. Commit each finished task. When a section works, commit it with a message that says what
and why. Small commits = easy recovery when something later goes wrong.
Prompts that work (copy these)
- "Read
skills/README.md, then load <skill> for this task."
- "Search the repo first — don't assume it's not implemented. Then propose a plan before coding."
- "Interview me before writing this: what do you need to know?"
- "Write this cell, then tell me exactly how I verify it worked."
- "Here is the error output, unedited: . Fix the actual cause, not the symptom."
Paste real errors whole — don't summarize them. The assistant fixes what it can see.
How to verify you're using this skill right
- Your conversations are short and single-purpose; you can name each one's task in five words.
- The assistant asked you at least one clarifying question before any multi-step work.
- Every finished section ran top to bottom before you committed it.
- You can explain, in your own words, everything you committed. If you can't — that's the signal
to slow down: core idea first, AI second.
