Treat Claude like a capable new colleague who takes you literally and sounds just as sure when it's wrong. Brief it well, ask where each claim came from, check the work before you trust it, and write down what it learns in your folder. The first five cards turn those four into prompts you can copy.
Two kinds of habit
Habits the labs build
You'll practice each of these in a lab. Keep the card open while you work.
Set up a session
Built in · Lab 1, then every labMost weak answers start before the first prompt. Open the folder for the job, check that Claude has read your AGENTS.md, start a fresh chat when the job is new, and brief it in four parts: the goal and why it matters, the files to use, what done looks like, and the limits.
Current models take instructions literally, and they do better when they're told why the work matters and what done looks like. The four parts give them both. The question at the top tells you right away whether Claude is working from your folder or from a blank page.
Lab 1 gives you a complete brief and your first AGENTS.md. Lab 2's brief has blanks to fill in, Labs 3 to 5 give you only questions under each part, and in Labs 6 and 7 you get just the four part names.The second chair
Built in · Lab 3, again in Labs 5 and 6Have Claude read your draft as a skeptic before anyone else does. It takes the role of a board member, a funder or a reviewer and tries to break the draft, and every catch is a problem you found first.
A wrong sentence from Claude reads just as smoothly as a right one, so a draft that looks finished can still carry mistakes. A separate critique pass, from a different role, puts more of them on the table while they're cheap to fix. Give it the files the draft came from, so it can check the numbers as well as the writing. It won't find everything, which is why the receipts rule and your own spot-check sit beside it.
Lab 3 builds it, has you settle its top three findings (fix it, reject it with a reason, or check the source), and saves it as a recipe. Lab 5 runs it as the meeting's skeptic on the narrative you write from your numbers, and at the end of room time Claude plays that skeptic on one of your slides in a fresh chat. Lab 6 turns it into a skill.The receipts rule
Built in · Lab 1, then every labEvery claim gets a source line, and Claude lists what it couldn't check, which is usually where the risk is.
Claude sounds the same whether it knows something or is guessing. A source line for each claim, plus an honest list of gaps, lets you tell the two apart before a board member or a funder does.
It's on your check-before-it-ships list from Lab 1, and in Lab 5 every number on a slide traces through the narrative to the rows it came from.Done means shown
Built in · every lab, most visibly Lab 4When Claude says it's finished, ask to see the work: the changed text, the actual number, the saved file. A report of success is a claim like any other.
An AI tool can report a step as complete when it didn't happen, or save a file that still holds the old version. A two-minute look at the real output catches it before anyone else sees it.
Every lab checks the output before it ships. Lab 4 makes the check visible with a change log that lists every edit to your sheet.Make the correction stick
Built in · Lab 1, then every labWhen you correct Claude, put the fix in your folder as well as the chat. A rule in AGENTS.md or your Org-Brain notes applies the next time the folder is used, while a correction typed into one chat mostly stays in that chat.
A fix in the moment works once. A rule in the folder works every time the folder is opened, by you, by your colleague, or by another tool you load it into. In Claude in the browser, ask for the line, paste it into AGENTS.md yourself and upload the new version to your Project.
AGENTS.md starts in Lab 1 and the Org-Brain folder in Lab 2. Every lab ends with "Save what you learned", where Claude suggests up to three lines for AGENTS.md from the day's fixes and adds only the ones you say yes to. In Lab 3 a real catch becomes a rule in your second chair, and in Lab 6 the second chair becomes a skill, a file that keeps its instructions for next time.Extras worth a minute
None of these belongs to a lab, so use them whenever they fit.
The hype filter
Extra · when a new tool shows upTell Claude who you are and what you're working on, then ask what in an article or video actually matters for that work. Context turns a generic recap into a decision about your week.
Without context you get a summary of the article. With it you get a judgment about your own work, and permission to skip the rest.
Options with consequences
Extra · any decisionAsk for options, what each one costs, and a recommendation. You keep the judgment, and Claude does the legwork of laying out the trade-offs.
A single answer hides the trade-offs and asks you to trust it. Options with their costs put the decision back with the person who knows the organization, and the habit works in any chat from the first week.
The handoff note
Extra · end of a long sessionBefore you close a long working session, have Claude write a note for whoever picks the work up next, and save the note in your folder.
Claude has memory now, and each Project keeps its own. Still, you can't easily read what it remembers, correct it or pass it to your colleague. A note in the folder is there for anyone who opens it, you included, which makes the folder the memory you can count on. In the browser, download the note and add it to your folder yourself.
The gap audit
Extra · before a meetingBefore a meeting, ask Claude what you don't know yet. The point is to find the open questions while there's still time to answer them.
Summarizing what you already know is easy and rarely helps much. A list of what you don't know, made before you walk into the room, gives you time to go and find it. The instruction to mark gaps as open keeps a guess from turning into a fact.
Stretch your usage
Extra · when your usage runs lowYour usage limit is one allowance shared across your chats and Cowork. Long conversations, large files, Research and multi-step tasks use it fastest. Four habits make it last.
- Start a fresh chat when the job changes, because an old chat drags every earlier job along with it.
- Carry the result forward. Ask for a short summary of what you worked out, then paste it into the new chat.
- Say what shape the answer should take, such as three bullets, one paragraph or a two-column table.
- Spot-check one claim yourself. Checking the most important fact against its source tells you more than asking Claude to review its own work.
A fresh chat that starts from a short summary does the same job without the weight of everything said before it. Files you keep in a Project count for less when they're reused, and a lower effort setting stretches your usage further on routine jobs.