Last verified 23 September 2026. Claude's free plan includes up to five projects, and each project keeps one set of saved instructions. A plain text file you copy the prompt from works in any assistant. Details for experts at the end of this book.

The idea in one page

A reusable prompt has four parts, followed by the input:

Three things decide whether it works.

One example beats ten adjectives. "Precise, consistent, professional" gives the model nothing to copy. One finished row shows the format, the units and the level of detail at once. Pick an example with a tricky case in it, such as a missing field, so the model also sees how you want gaps handled.

Change one thing, then look. When the output is wrong, change one part of the prompt, run it again on the same five inputs and compare. Keep a short log: R1: added an example, and all five rows now use the same format. It's the only way to know what worked, and it stops you from breaking a part that was already fine.

Missing information can't be fixed with wording. Before you touch the prompt, ask one question: is the answer actually in the input? If an abstract doesn't state a sample size, no instruction will produce the right number. The best a prompt can do is make the model say so. The same goes for the role: "You are an expert" adds no knowledge and no authority. So you have two repairs, and they don't mix:

• An unclear instruction means the answer is in the input but comes out wrong. Change the prompt.
• A missing fact means the answer isn't in the input. Get it from somewhere else, or accept "not stated".

When the fact lives in another document, the fix is to hand the model that document as well. T03-L02 · Build a Context Packet Your Team Can Rerun shows how to pack the right documents into the chat, and T04-L01 · Ask Your Own Documents and Check Every Citation shows tools that search your documents for you.

Worked example: a Friday extraction prompt

The prompts are real templates you can copy. The results described are typical examples written for this book.

Lab: five abstracts into the literature table

Structure it. "Each Friday I add about five new abstracts to our group's literature table. I need four details per abstract, in the same format every time, so I can copy them straight in as rows. A missing detail must show as missing, never guessed."

Understand it. Jonas, a postdoc who keeps his group's table, pastes one abstract and asks: "Extract organism, sample size, method and main finding." That works. He pastes the other four in the same chat. The answers come back in three different formats. The sample size appears as "24 mice", "n=24" and "24 animals (12 per group)". For abstract 4, which only says "a cohort of mice", the answer gives "approximately 20". He learns two different things. The format problem is his: he never said what a row should look like. "Approximately 20" is the other kind of problem: that fact isn't in the abstract, so no wording will make the answer right. Back to Structure it: "…sample size as the total number only, or 'not stated'."

Choose tools. Five abstracts a week is too few to automate. A saved prompt in his chat assistant is enough, kept as the instructions of a Claude project. If it grows to fifty a week, the next step is an automation, a small set-up that runs the prompt for him without pasting (T05-L01 · Your First Automation: Build It, Break It, Switch It Off).

Build it. His first full prompt has the role, the task and the constraints, but no example yet. He types it into a new chat, pastes the five abstracts where the square-bracket line is, and sends it all as one message. The lines in angle brackets only mark where the abstracts start and end:

On the same five abstracts, the format is now identical and abstract 4 says "not stated". Two details still vary: how the method is worded and how the organism is named. Then come three revisions, one change each. For each one he opens a new chat, so the earlier answers don't steer the new one, and pastes the changed prompt with the same five abstracts:

• R1, add an example. He adds one finished row, for a zebrafish study: "Danio rerio | 36 | fin regeneration assay | Compound X reduced fin regeneration by 40%." The method column changes from long phrases to short labels like the one in the example.
• R2, one constraint on the finding. Two findings had borrowed the authors' interpretation from the conclusion ("suggesting a new therapeutic target"). He adds: "Main finding: the measured result only, not the authors' interpretation." Now five of five findings are results.
• R3, one constraint on the organism. "Mice", "mouse" and "Mus musculus" (the Latin name of the house mouse) all appear. He adds: "Organism: species name in Latin." All five now match.

For abstract 4 he opens the full paper and takes the sample size from the methods section himself. That was never a prompt problem. He saves the final prompt as the project instructions, with the three-line log in his text file. Next Friday he pastes five new abstracts and checks five rows.

Company: supplier invoices into the payments sheet

Structure it. "Every Friday I enter about eight supplier invoices into the payments sheet. A useful result is eight rows I only have to check, not type."

Understand it. Sara, who enters the invoices in accounts, pastes one and asks for "the invoice details". She gets a friendly paragraph that includes the bank details and the VAT breakdown, with the date written as "12/09/2026", which could be 12 September or 9 December. For an invoice that says "payable within 30 days" and prints no due date, the chatbot works one out. The date format is a prompt problem. The due date is a missing fact: the invoice gives a payment term, not a date, and working out dates is exactly the kind of arithmetic a chatbot shouldn't do for her (T01-L01 · From Idea to Working Result: The Four-Quadrant Method). Back to Structure it: "…the due date only if it's printed. Otherwise the payment term as written, and the sheet calculates the date."

Choose tools. Invoices carry bank details and sometimes names, so she uses the assistant her company approved. It's a saved prompt, and there's no automation yet.

Build it. Her prompt has the same parts as the lab one: a role, a task (supplier, invoice number, invoice date, gross amount with currency, due date or payment term), constraints, and then the pasted invoices. Then three revisions, each in a new chat on the same eight invoices:

• R1, add an example row with the date as 2026-09-12. Every date now comes out in one format.
• R2, one constraint on the amount: "Gross amount including VAT, with the currency code." Two invoices had given the net amount. Now none do.
• R3, one constraint on the due date: "Due date only if printed. Otherwise write the payment term as it appears, for example 'net 30'." No more calculated dates. Her spreadsheet works out the date from the invoice date.

She checks each row against its invoice, the same way every week. It takes her a few minutes instead of the half hour of typing.

In both cases the example fixed the format and the constraints fixed the content. The one detail no prompt could fix was the one that wasn't in the input.

What goes wrong

Politeness instead of rules. Symptom: "please be very careful and accurate", and the same errors come back. Fix: replace every adjective with a rule or an example.

Asking for "professional" or "consistent" without showing it. Symptom: each run has a different idea of what that means. Fix: add one finished example of the output.

Changing five things at once. Symptom: the output improved and you can't say why, or it got worse and you can't find the cause. Fix: one change per revision, one line in the log.

Blaming the prompt for missing information. Symptom: you reword the prompt five times for a detail the input doesn't contain. Fix: check the input first. Add a "not stated" rule and get the fact from somewhere else.

Testing on one input. Symptom: the prompt works on the abstract you wrote it for and fails on the next one. Fix: always run it on the same five inputs, including the awkward one.

A prompt that works once and is never saved. Symptom: next Friday you write it again from memory, slightly differently. Fix: keep it as project instructions or in a text file, with the log at the top.

Do it yourself (25 minutes)

1. Structure it (4 min). Pick a task where you pull the same details out of documents again and again. List the details you need, the format you want, and what counts as missing. Collect five inputs, including one awkward one.
2. Understand it (5 min). Ask for the details in a plain request on all five inputs. List every problem. For each one, write instruction or missing fact.
3. Choose tools (2 min). Decide where the prompt will live: in a Claude project or in a text file. Check that you're allowed to paste these inputs into this assistant.
4. Build it (12 min). Write the four-part prompt in your text file. Open a new chat, paste the prompt with all five inputs, and send it. Then make three revisions, one change each. Run each one in a new chat on the same five inputs. Write one line per revision: what you changed, what it did.
5. Build it (2 min). Save the final prompt. In Claude, create a project and paste the prompt as its instructions (the clicks are in Details for experts). Otherwise keep it at the top of your text file, above the log.

Exit check

What you hand in: one saved prompt, its three revisions, and a one-line note for each saying what it changed on the same five inputs, plus the final output.

It passes when a colleague can see what each change did, and every detail missing from an input reads "not stated" instead of a guess.

Rule to remember

Change one thing, then look.

Details for experts

Checked on 23 September 2026. You don't need this to follow the book; it's for setting things up or answering a colleague.

Further reading and tools

Next books

• T03-L02 · Build a Context Packet Your Team Can Rerun: when the prompt is fine but the model never had the document that mattered.
• T03-L03 · Put a Test Set Under the Prompt Your Team Depends On: a proper test set, so "better" is a score, not a feeling.
• T05-L01 · Your First Automation: Build It, Break It, Switch It Off: when the weekly batch is too big to paste.
• T07-L02 · Hand a Task to an Agent, and Check What It Did: the same job description, given to an agent that reads and writes files.
• T12-L01 · Decide What You May Paste Into AI at Work: what you may paste, and where.

Tools

• Claude: Taught. Projects hold the prompt on the free plan.
• ChatGPT, Gemini: Catalogued. The same prompt works in either.