Brief it like a new hire

The Lean job: give someone the context to do the work

Every Lean leader has watched a new person get a job wrong because nobody said who the work is for, what good looks like, or what to do when something doesn't fit. The fix was never "try harder". It was to put what you know into the instruction. An assistant has the same problem, with two differences. It cannot walk over and ask what you meant unless you invite it to. And it tends to agree with you and to guess when the facts run out, as Lesson 1 showed.

A briefing closes the gaps before the work starts.

Predict first

Here are two ways to ask for the same thing. Both come with the same ten scrubbed stops from Lesson 2.

A: Summarize these stops.

B: You are helping a packaging team lead prepare a five-minute Monday huddle. Summarize last week's stops, using only the table below. The huddle ends with the team choosing one thing to try, so end with one question for them. Readers are line operators: plain words. Format: three bullets under 20 words each, then the question. Use only the minutes and reasons in the table, and show your arithmetic. If a cause isn't in the table, write [NEEDS GEMBA]. If you can't tell, say I don't know.

Before you read on, write down two ways the answers will differ, and which one you would trust more. Then think about what B had to know to write itself, and who had to supply that.

What goes in a briefing

Anthropic's own guide to prompting says to be specific, say why, state the output format, give three to five examples, put long documents first, and set a role. Here that becomes six questions you answer before you send:

Why the last one is worth a line

Lesson 1 said models tend to fill a gap with a guess, and that training and scoring can reward guessing over "I don't know". Anthropic's guidance on reducing made-up answers says to allow the assistant to say it doesn't know, to have it quote the text it relies on, and to check by running it again. One line costs nothing. It is not a guarantee, so you still read the answer, but it can move the default.

Put long documents first

If you paste a long SOP or a long log, Anthropic's guidance is to put it before your question. Separately, as Lesson 1 explained, facts in the middle of a long input are used less reliably.

A briefing is a paste

Everything in a briefing leaves the building too. Apply Lesson 2 to it: no names, no formulas, no customers. If a briefing needs your plant's terms, put the terms in and leave the people out.

What matters, and what has little evidence behind it

• Context and examples matter. They are what the assistant can't get anywhere else.
• Format can matter a lot, in odd ways. In one study of an older open model, changing only how a prompt was laid out moved accuracy by up to 76 points. Test your briefing on your own cases, and don't assume a layout that worked once will keep working.
• A role is cheap and unreliable. Studies of personas such as "you are an expert" found they did not reliably make factual answers more accurate. Setting a role is fine. Don't count on it to fix a thin briefing.
• "Think step by step" is worth less than it was. A 2025 preprint found its gains shrinking as models improved. For arithmetic, a method that shows its work or runs code is better, and Lesson 5 takes that up.

Where a briefing lives

If you retype the same briefing every time, you are doing work that standard work exists to remove. There are six homes for context, and they differ in how often the assistant sees them and who controls them.

How to choose

• The same words every time you work on one line or one problem: a Project.
• A preference that applies to everything you do: custom instructions.
• A procedure the team runs the same way, such as the 12-step check: a skill. You write one in Lesson 7.
• A procedure your team must follow: never memory. It is inferred automatically and isn't controlled content, so it has no named owner, version or review date.

Your turn: the one-liner against the briefing

About 25 minutes, using the same ten scrubbed rows from Lesson 2. They are repeated here so you can start without going back.

Day, Shift, Customer, Minutes, Reason
Day 1, Day, C1, 12, film splice, operator re-threaded
Day 1, Swing, C2, 25, label jam
Day 2, Day, C1, 8, film splice
Day 2, Swing, C3, 31, waiting on forklift, asked lead
Day 3, Day, C2, 15, label jam
Day 3, Night, C3, 40, changeover
Day 4, Swing, C2, 22, label jam
Day 4, Night, C1, 18, film splice
Day 5, Day, C3, 35, changeover, operator waiting on kit
Day 5, Swing, C2, 20, label jam

1. Run A (4 minutes). In a new chat, paste the rows and send only:

   Summarize these stops.

   Save the answer.
2. Write your briefing (8 minutes). Use the six boxes on the worksheet. Here is a model to adapt:

   You are helping a packaging team lead prepare a five-minute Monday huddle. Summarize last week's stops, using only the table I paste after this. The huddle ends with the team choosing one thing to try, so end with one question for them. Readers are line operators: plain words. Format: three bullets under 20 words each, then the question. Use only the minutes and reasons in the table, and show your arithmetic. If a cause isn't in the table, write [NEEDS GEMBA]. If you can't tell, say I don't know.
   Examples of a good bullet: "Seal-bar resets: 41 of 190 minutes, the biggest share." and "Hopper refills: 3 stops, 55 minutes."
   Example of a good question: "What is different on the shift where it happens most?"

3. Run B (4 minutes). In a new chat, send your briefing, then the rows. Save the answer.
4. Compare (6 minutes). Did it follow the format? Which bullets come from the table? Add the minutes by reason yourself and check its numbers. Did it name any cause that is not in the table? Did anything get marked [NEEDS GEMBA] or "I don't know"?
5. Move it (3 minutes). Put the reader, the format, the examples and the "I don't know" line into a Project's instructions. Open a fresh chat in the project and paste only the rows. Is the answer close to B's? One instruction can change a lot, and it still isn't a guarantee, so keep checking.

Knowledge check

Your task: write one briefing you will reuse

1. Open the printable worksheet and fill in the six boxes for one job you do every week.
2. Decide where it lives: a Project, custom instructions or, later, a skill. Write down who owns it and when you will review it.
3. Use it three times this week. After each, note one thing the briefing should have said, and add it.

Open the printable worksheet

In the toolkit

A job breakdown sheet is the Lean way to brief a person on a job: the important steps, the key points and the reasons. The toolkit's TWI Job Breakdown skill builds one, with a training timetable. Read it next to your own briefing. It is the same job, for a person.

Sources

Where the facts in this lesson come from. Facts last checked October 1, 2026. Plans, menus and product names change quickly, so check the vendor's own page. If you find a fact that is out of date or wrong, tell me.

• Anthropic: Prompting best practices
• Anthropic: Reduce hallucinations
• Zheng et al. (EMNLP Findings 2024): how personas affect accuracy
• Sclar et al. (ICLR 2024): sensitivity to prompt format
• Meincke et al. (2025 preprints on prompting): first report and second report
• Why Language Models Hallucinate, Kalai et al. (preprint)
• Claude Help: What are Projects, What are skills and How does Claude's memory work
• Anthropic: Agent Skills overview and How Claude remembers your project
• Google Gemini Help: Transition from Gems to skills