A free primer from Albino Geek Services. For everyone who uses Claude, ChatGPT, Cursor, or Codex and is tired of answers that are confidently almost right.
Your assistant is not lazy and it is not psychic. Every time it lacks a fact, it fills the gap with the most likely guess, and the most likely guess is the average of everyone else's situation. These ten rules close the gaps before it has to guess. None of them needs code.
1. Say what done looks like
"Help me with this email" gives the assistant nothing to aim at. "Write a reply under 120 words that declines the meeting, offers two times next week, and keeps the tone warm" gives it a finish line.
Before you send a request, finish this sentence: I will know this worked when... Then paste that sentence into the request. Length, format, audience, and the one thing it must include are the usual missing pieces.
2. Brief it like a new hire
Imagine handing the task to a sharp temp on their first morning. They are capable, but they have never met your client, do not know your price list, and have no idea the last proposal went badly.
Whatever you would tell that temp, tell the assistant: who the reader is, what has already been decided, and what went wrong last time. Two or three sentences of background fix more bad answers than any clever wording.
3. Paste it, do not describe it
"The policy says we can't refund after 30 days, roughly" invites the assistant to invent the rest of the policy. Paste the policy.
The same goes for the email you are replying to, the spreadsheet columns, the style guide, the error message. The assistant works from what is in front of it. Your summary of a document is a lossy copy, and it will fill the holes with plausible fiction.
4. Name what it cannot guess
Some facts no amount of intelligence can infer: your budget, your deadline, the competitor you never mention by name, the word your boss hates, the fact that your audience reads on a phone.
Keep a short list of these for each kind of task you repeat. Constraints are cheap to state and expensive to discover in the third draft.
5. Show one example of good
If you have a past email, report, or post you liked, paste it and say "match this tone and structure". One real example beats a paragraph of adjectives like "professional but friendly", which every reader interprets differently.
Say which parts to copy (structure, length, tone) and which parts not to (the facts, the names). Otherwise you may get your old example back with the dates changed.
6. Give it permission to say "I don't know"
Assistants are trained to be helpful, and an invented answer looks helpful right up until someone checks it. Add one line to requests where accuracy matters:
If you are not sure of a fact, say so and tell me what you would need to confirm it. Do not fill gaps with a best guess.
This single sentence turns a confident wrong answer into a short list of things to look up.
7. Ask for its assumptions first
For anything bigger than a paragraph, ask the assistant to start with the assumptions it is making, then stop and wait:
Before you write anything, list the assumptions you would have to make to do this. Then wait for me.
You will usually spot the wrong one in ten seconds. Correcting an assumption costs one reply; correcting a finished draft built on it costs the whole draft.
8. One job per message
"Summarize this contract, flag the risky clauses, draft a reply to the vendor, and suggest a negotiation strategy" is four jobs. The assistant will do all four at half attention, and you will not be able to tell which part went wrong.
Send them one at a time. Check each result before the next step depends on it. It feels slower and finishes sooner.
9. Start fresh when a chat goes stale
A long conversation is a pile of everything said so far, including the draft you rejected and the instruction you changed your mind about. Past a certain length, the assistant starts blending old and new.
When answers start drifting back toward something you already corrected, open a new chat. Paste a short summary of what was decided and carry on. The assistant forgets nothing on purpose; it just cannot tell which of your instructions are still current.
10. Fixed it twice? Make it a rule
If you have corrected the same thing in two separate chats ("stop using exclamation marks", "we are Canadian, use CAD", "never promise a delivery date"), that correction belongs in a rule the assistant reads every time, not in your next message.
Every major assistant has a place for standing instructions:
| Assistant | Where standing rules go |
|---|---|
| Claude app | Settings, then Instructions for Claude (all chats), or a Project's instructions (one project) |
| ChatGPT | Settings, then Personalization, then Custom instructions (all chats), or a Project's settings (one project) |
| Cursor | Customize, then Rules (all projects), or an AGENTS.md file in the project |
| Claude Code | A CLAUDE.md file, in your home folder's .claude directory or in the project |
| Codex | An AGENTS.md file, in ~/.codex/ or in the project |
Keep each rule short, specific, and checkable. "Write clearly" is a wish. "Keep paragraphs under four sentences" is a rule, because you can see at a glance whether it was followed.
Where to go next
These ten rules cover the single conversation. The AI Field Guide covers the rest: how assistants use context, how to read what the model did and catch it when it drifts, how to write rules and reusable skills, the red flags that mean an answer needs checking, and step-by-step setup for Claude, ChatGPT, Cursor, and Codex.
If you would rather start from finished files than blank pages, the Skills & Rules Starter Kit gives you fill-in rules files and six ready skill templates, each with a guide to adapting it to your work.
Albino Geek Services · albinogeek.com