This mini-book is for Level 1 User. You are improving one task for your own work. You remain responsible for choosing permitted material, checking the result, and deciding whether it is useful.
2. A longer vague request is still vague
You ask a chatbot to “summarise these notes.” It returns a generic paragraph that could describe almost anything. You try again with more words. The second answer is longer, more polished, and equally useless.
The problem is not a missing magic phrase. The chatbot had to guess what the source was, who needed the result, what useful length meant, and which details mattered. Adding enthusiasm or politeness did not remove those decisions.
A useful request supplies the work instead of describing it, names the reader and length, shows a small example when format matters, and states what must be excluded. You then revise one condition at a time. That gives you a better answer and teaches you which instruction caused the improvement.
3. After this you can
- Provide source material, audience, length, example, and exclusions in one request.
- Improve a prompt through three named revisions instead of restarting blindly.
- Explain what changed after each revision.
- Choose whether the task primarily needs source search, drafting, document work, or structured analysis.
4. Prerequisites
T01-L01· What AI can and can't do for your work.- One short public, synthetic, or explicitly approved source.
- An approved chatbot, or the supplied example outputs if you prefer to work offline.
- A text file in which you can keep prompt versions and results.
Do not upload personal data, confidential company material, unpublished research, credentials, contracts, customer records, or production exports to an unapproved service.
5. The idea in one page
Use four moves in this order.
1. Give the material
Attach or paste the permitted source instead of describing what it probably contains. If the task depends on facts, draw a visible boundary around the source and say what to do when information is missing.
Use only the material inside <source>. If a requested detail is absent,
write: not stated in the source.
This does not make the answer true. It makes unsupported additions easier to detect.
2. Name the reader and length
“Make it clear” leaves the model to invent the audience and level of detail. Replace it with an observable target: “Write for a new lab member in 120–150 words” or “Write three bullets for a client who has not seen the project.”
3. Show one small example
Use an example when the required pattern is difficult to describe. The example should demonstrate shape, tone, or classification, not introduce unsupported facts. One representative example is often enough for a simple task.
Pattern
Finding: [one supported sentence]
Evidence: [exact phrase from the source]
Open point: [missing information or none]
4. State what to leave out
Name the common failure before it happens: no invented citations, no recommendations, no customer promises, no interpretation beyond the source, or no technical terms that the reader would need explained.
Change one thing at a time
Label your versions V1, V2, and V3. Keep the task and source fixed. Change one condition, compare the output, and write one sentence about the effect. If you rewrite the source, audience, format, and constraints together, you cannot tell which change helped.
Choose a tool category before a brand
| Job | Starting point to compare | Check before use |
|---|---|---|
| Find current public sources | Perplexity or another search-with-sources tool | Open the original sources and verify dates and claims. |
| Draft carefully from supplied text | Claude or another tool with document handling | Confirm file support, data policy, and whether the answer can point back to the supplied text. |
| Inspect a table | ChatGPT or another tool with structured data analysis | Recalculate totals and inspect changed rows. |
| Work inside an office suite | Gemini or another approved workspace integration | Confirm permissions, connected content, sharing, and organisational policy. |
Current-capability check — last verified 4 September 2026. The provider categories above are starting points, not exclusive recommendations. Interfaces, plans, file limits, analysis features, and data controls change. Confirm each capability in the current primary documentation and your account before choosing it. Select for the job and approval boundary, not for a familiar brand.
6. The worked example: three visible revisions
Lab framing: rough notes to a methods paragraph
Use this synthetic source:
<source>
Fixture: Northstar assay demo
Material: four synthetic samples S-01 to S-04
Preparation: mix for 30 seconds
Measurement: record two readings per sample
Reporting rule: report the mean; if readings differ by more than 5%, write review needed
Temperature: not stated
</source>
Baseline: vague request
Write this up as a methods paragraph.
Illustrative output:
Four synthetic samples were prepared and measured under standard conditions.
Two readings were taken for each sample and averaged to provide reliable results.
Samples with inconsistent readings were reviewed before reporting.
The fluent output invents “standard conditions,” weakens the exact 5% rule into “inconsistent,” and never reveals that temperature is absent. The request gives no reader or test for success.
Revision 1: add one output target
Write this up in 35-55 words for a new lab member.
Illustrative output:
Prepare four synthetic samples, S-01 to S-04, by mixing each for 30 seconds.
Record two measurements per sample and calculate the mean. Review samples whose
readings are inconsistent. The samples are measured under standard laboratory
conditions before the results are reported.
Effect: the audience and length are now observable, but the same unsupported condition and weakened threshold remain. Only the output target changed.
Revision 2: add one evidence boundary
Keep Revision 1. Use only <source>. If a detail is absent, write "not stated".
Illustrative output:
Four synthetic samples, S-01 to S-04, are mixed for 30 seconds. Two readings are
recorded for each sample and their mean is reported. If the two readings differ
by more than 5%, the result is marked “review needed.” Temperature is not stated
in the source.
Effect: the unsupported condition disappears and the exact threshold survives. Only the evidence boundary changed.
Revision 3: add one example pattern
Keep Revision 2. Follow this pattern:
Preparation: [supported sentence]
Measurement: [supported sentence]
Exception: [exact rule]
Missing condition: [detail or "none"]
Illustrative output:
Preparation: Four synthetic samples, S-01 to S-04, are mixed for 30 seconds.
Measurement: Two readings are recorded per sample and their mean is reported.
Exception: If readings differ by more than 5%, write “review needed.”
Missing condition: Temperature is not stated.
Effect: the facts do not change; their shape becomes easier for a new lab member to inspect. Mira checks each line against the fixture before private reuse.
Company framing: rough notes to a client-ready summary
Use this synthetic source:
<source>
Project: Harborlight demo
Completed: draft navigation and two test pages
Open: mobile review and accessibility check
Decision needed: approve one of two fictional colour directions
Delivery date: not stated
Commercial terms: not stated
</source>
Baseline: vague request
Create a polished client update.
Illustrative output:
The Harborlight project is progressing well. The navigation and test pages are
complete, and the remaining mobile checks will be finished before delivery.
We recommend proceeding with the stronger colour direction.
The output invents progress, a delivery expectation, completion, and a recommendation. “Polished” does not define a factual boundary.
Revision 1: add one output target
Write three brief bullets for an internal account reviewer.
Illustrative output:
- Progress: Navigation and two test pages are complete.
- Next: Mobile and accessibility reviews will be completed before delivery.
- Recommendation: Approve the preferred colour direction.
Effect: the destination and length are clear, but unsupported completion, delivery, and recommendation language remain. Only the output target changed.
Revision 2: add one evidence boundary
Keep Revision 1. Use only <source>. Mark missing information "not stated".
Illustrative output:
- Completed: Draft navigation and two test pages.
- Open: Mobile review and accessibility check.
- Decision needed: Approve one of two fictional colour directions.
Effect: the promises and recommendations disappear. Only the evidence boundary changed.
Revision 3: add one example pattern
Keep Revision 2. Follow this pattern:
- Completed: [supported item]
- Open: [supported check]
- Your decision: [supported choice]
Illustrative output:
- Completed: Draft navigation and two test pages.
- Open: Mobile review and accessibility check.
- Your decision: Choose one of two fictional colour directions.
Effect: the final shape matches the reviewer's job without changing any fact. Jonas verifies each bullet and keeps the exercise output private; a responsible colleague would send any real update.
The important practice in both versions is the same: preserve the source, save each prompt, change one instruction, and record the observed effect.
7. What goes wrong
You describe the source instead of supplying it
Symptom: the answer fills gaps with generic or outdated information.
Fix: attach the approved document or paste the permitted excerpt. Ask for not stated when evidence is absent.
“Summarise” has no reader
Symptom: the output has the wrong detail, tone, or terminology.
Fix: name the reader, purpose, length, and destination.
Every revision changes everything
Symptom: V3 is better, but you do not know why and cannot repeat the result.
Fix: freeze the source and task. Change one audience, format, example, or exclusion at a time.
Search and drafting are confused
Symptom: a search-oriented answer becomes the final prose, or a drafting tool is expected to discover current evidence without a search step.
Fix: separate finding sources, reading sources, drafting, and verification. Use the capability appropriate to each step.
Fluency ends the review
Symptom: the first polished answer is accepted without checking the source.
Fix: apply T01-L02: inspect names, numbers, dates, claims, and omissions before the output leaves your desk.
8. Do it yourself: three revisions in 25 minutes
Minutes 0–5: choose one short approved source and one concrete transformation. Save the source unchanged.
Minutes 5–9: write and run V1. Capture the output without silently correcting it. Mark one visible failure.
Minutes 9–15: create V2 by adding only one output target, such as a named reader and observable format. Compare it with V1 and record the effect in one sentence.
Minutes 15–21: create V3 by adding only one evidence boundary or one example pattern. Keep everything else fixed. Record the effect.
Minutes 21–25: compare the final output with the source. Keep the smallest change that fixed the named failure and remove instructions that did not affect the result.
9. Exit check
Deliver exactly one artifact: a before/after pair containing the original prompt and output, the improved prompt and output, and one sentence naming the single change that fixed the selected failure.
It passes when another person can see the same source boundary, identify the changed instruction, and confirm that the claimed improvement appears in the output without adding unsupported facts.
10. Rule to remember
Attach the thing. Then say who it is for.
11. Further reading & tools
Curriculum links:
- Taught:
T01-L02· Trust but verify — checks the output after prompting. - Catalogued:
T03-L01· Prompting basics — turns this four-move habit into a reusable prompt structure. - Catalogued:
T02-L02· Your daily driver: files, projects and memory — continues from single chats to bounded workspaces, files, and memory.
Tools taught in this book:
- Taught: Claude — Anthropic prompting best practices (opens in a new tab) — for careful drafting from supplied text.
- Taught: ChatGPT — OpenAI data analysis help (opens in a new tab) — for inspecting uploaded structured data.
- Taught: Gemini — Google Workspace help for Gemini in Sheets (opens in a new tab) — for work inside an approved Google Workspace.
- Taught: Perplexity — Perplexity Search documentation (opens in a new tab) — for finding current public sources before opening and checking them.