Skip to lesson
Lesson 05 / 2020 min readUp to date as of 19 October 2026

How to write useful prompts

Goal, sources and checking the answer

Say what you need, which data to use and how you will judge the result. Start with the short template; use the detailed structure for harder tasks.

Reading time is approximate. Allow extra time for the exercises; practice estimates are shown with the workshops.

After this lesson
  • Write a prompt with a clear output and quality criteria
  • Specify sources, limits and a stopping condition
  • Compare two answers against the same criteria
Section being translated

The chrome and course program are available in English. Deep training modules (prompt library, AI glossary, playbooks) are being translated - some content below is still in Polish.

Edition date and feature availability

Material revised: 19 October 2026; The interface and folder guide was revised on 19 October using supplied screenshots and OpenAI’s public plugin repositories; this does not reverify product access. The models in Chat, Work and Codex were checked on 16 October 2026, and the GPT-5.5 retirement date and Word access rules on 14 October 2026; other prices and availability were not fully reverified. Check the documentation and your account before using a feature: access depends on plan, region and administrator settings. Use the examples for practice and review the results yourself.

See how it works

Anatomy of an assessable brief

  1. Goal

    Which decision or action should become easier?

  2. Materials

    What supports the result? Specify files and scope.

  3. Result

    Define format, audience and length.

  4. Limits and review

    Define restrictions, missing data and acceptance criteria.

A concrete example

‘Compare these two offers in a table: scope, cost, risk. Use only the files. Mark missing prices as “no data”. Link every row to its source.’

What if the file contains no price?

Mark missing data and request it. Do not guess the price.

Course-data workshop · 20–30 min practice

See what a clear request changes

A manager asks whether sales increased. The file contains a duplicate, a draft proposal, a foreign currency and an adjustment. Prepare an answer you can verify. All companies and transactions are fictional; no CRM or email is needed.

Lessons 05, 06 and 16 use the same example. Keep your earlier result and build on it rather than starting over. Time is an estimate.

Download CSV dataDownload instructions
  1. First ask: “Summarise sales from this file.” Record its assumptions. A vague request is not necessarily wrong — the model may ask for rules.
  2. Use explicit counting rules. A good result includes a table, sources for the numbers and exclusions, not just “sales increased”.
  3. Calculate the totals independently with a calculator or spreadsheet. Check the duplicate, currency, status and adjustment sign before reviewing the commentary.
Request with rules and the complete data
Prepare a report for the sales manager: a two-period table, difference in PLN, percentage change, exclusions with IDs, and up to three sentences of commentary. Use only the data below. Choose available calculation tools; show the formula and documents included in each total. Do not invent causes of growth. Ask for missing data. Do not publish or send anything.

Compare 1–7 with 8–14 October 2026 (inclusive). Sum posted invoices and net adjustments in PLN. Count identical duplicate documents once; flag conflicting duplicates for clarification. Exclude draft proposals. Report EUR separately without conversion. Include the negative adjustment. Change % = (new total − old total) / old total × 100%; do not calculate an ordinary growth percentage when the old total is zero.

document_id,date,customer,document_type,status,currency,net_amount
FV-001,2026-10-01,Company A,invoice,posted,PLN,1000
FV-002,2026-10-02,Company B,invoice,posted,PLN,1500
FV-003,2026-10-03,Company C,invoice,posted,PLN,500
FV-004,2026-10-08,Company A,invoice,posted,PLN,1200
FV-005,2026-10-09,Company B,invoice,posted,PLN,1800
FV-005,2026-10-09,Company B,invoice,posted,PLN,1800
FV-006,2026-10-10,Company D,invoice,posted,PLN,600
OF-007,2026-10-10,Company E,proposal,draft,PLN,900
FV-008,2026-10-11,Company F,invoice,posted,EUR,400
KOR-009,2026-10-11,Company A,adjustment,posted,PLN,-200

Find the error before sharing the report

“Sales rose from PLN 3000 to PLN 6500, or 116.67%. The marketing campaign worked, so increase its budget by 20%.”

A deliberately incorrect training example, not an observed output from a particular model. Which mistakes concern the inputs, and which are unsupported conclusions?

Check the answer and why it is correct
Net value under the exercise rules
PeriodDocumentsPLN
1–7.10.2026FV-001, FV-002, FV-0033000
8–14.10.2026FV-004, FV-005, FV-006, KOR-0093400

Swipe the table sideways to see the amounts.

Difference: PLN 400. Change: 400 / 3000 × 100% = 13.33%.

6500 adds a duplicate 1800, draft 900 and EUR amount 400 to the correct 3400. A correct formula does not fix incorrect input selection. The −200 PLN adjustment is already included in 3400.

The file contains no campaign data. You can describe the change, but not its cause or the appropriate marketing budget.

Download answer and assessment rubric

Now try without the ready-made prompt

Change FV-006 from PLN 600 to PLN 800. Predict the change, then request an update in your own words. Save your report and the mistakes you spotted.

Second-trial answer

Second-period total PLN 3600, difference PLN 600, growth 20%. Exclusions are unchanged. If you leave the amount blank instead, the known portion is PLN 2800, but the full total needs the missing data.

Part 05 - Prompting GPT-5.6 and GPT-6

Start with the result, that you need

In shortDescribe the output, allowed sources and how to check it. For example: a short follow-up email based on meeting notes, with a suggested next date.

Try the short template below. For tasks with several tools or decisions, open the eight-section structure and additional tips. API cost depends on token and tool usage.

Five pillars of a good prompt

Instead of scripting every step - describe the goal and let the model choose the path.

What the result must be

Describe the end state - not every step to get there.

Success criteria

How will you know the task was done well?

Key constraints

What the model must NOT do, change, or add.

Allowed sources

Which data, files, links it can use.

When to stop

Stop rule - when work is done, when to ask a question.

prompt.txt - minimal template
Produce [concrete result].
Use [allowed sources]; flag missing information instead of guessing.

I care about [top quality criteria].

Preserve [things that must not change].

Do not add [what the model must avoid].

Before finishing, verify [quality check].
💡GPT-5.6 and GPT-6 golden rule: describe the output and the checks. Let the model choose the sequence where your procedure does not require a particular order. Review the result before using it.

Same goal - two prompts

In this example, a follow-up is a message after a meeting. CTA means a call to action, such as asking the client to choose a date.

Old style - micromanagement

"Write an email to the client. First greet them. Then thank them for the meeting. Then describe the offer in 3 bullets. Always be polite. Never mention the price without a discount. You must end with a call to action…"

GPT-5.6 / GPT-6 style - goal + criteria

"Write a follow-up email after a client meeting [context]. Goal: book another call this week. Good version: short (under 120 words), concrete, includes 1 client value and 1 time slot. Keep our tone - direct, no corporate. Don't mention pricing. At the end check that the CTA is clear."

More detailDetailed structure and 11 tipsEight sections can help with complex tasks. You do not need them for every short request.

Structure of a complex prompt

8 sections per OpenAI

That doesn't mean every prompt must be a document. But for harder tasks it's worth walking through this list.

  1. 1
    Role
    Who the model is for this task
  2. 2
    Personality
    Tone, communication style
  3. 3
    Goal
    The concrete result to achieve
  4. 4
    Success criteria
    How to know the output is good
  5. 5
    Constraints
    What is not allowed / to avoid
  6. 6
    Tools
    What it can use, what it has access to
  7. 7
    Output
    Format and structure of the answer
  8. 8
    Stop rules
    When to end, when to ask

11 specific OpenAI recommendations

01

Cut repetition

Don't restate the same rule five ways and don't add examples that change nothing. A longer prompt ≠ a better prompt.

Evidence: Internal OpenAI tests: simpler instructions = +10-15% quality and -41-66% token usage.

02

Result, not steps

Instead of "first X, then Y, then Z" write "Solve end to end. Good execution means…". GPT-5.6 picks the path better than older models.

03

State quality criteria

Say explicitly what MUST be in the answer: decision, rationale, risks, next step. Stronger than yet another instruction paragraph.

04

Careful with "always / never / must"

Reserve them for invariants: safety, format, bans. Elsewhere give decision criteria and let the model read context.

05

Stop rules

Say when to search more, when it's enough, when to ask, and when to state "data missing" instead of guessing. Without this the model loops.

06

Tone ≠ collaboration style

Tone: direct, formal, ironic. Style: asks vs. researches, flags uncertainty or not. Two separate axes - set both.

07

Set autonomy level

"Don't ask about small stuff; assume reasonable defaults and flag them. Only ask when a key detail is missing." - often enough.

08

Say what to KEEP

Not "improve this text" but "keep meaning, facts, length and tone - improve fluency and clarity, add nothing new".

09

Rules for sources

Which facts need confirmation, how to flag guesses, what to do with contradictions, when to say "no data" instead of guessing.

10

Ask it to check the result

Code → tests. Graphics → layout and legibility. Text → does it answer everything and skip nothing? A simple self-check lifts quality.

11

More "thinking" won't fix a weak prompt

Before cranking reasoning, check: goal clear? criteria present? no contradictions? model knows how to verify? Usually the issue is here, not model power.

What changes with GPT-6

Shorter prompts

GPT-6 follows the goal and criteria better. Cut step-by-step lists and repeated warnings - they cost tokens and can confuse the model.

Model in the prompt? No

You choose the model in the picker, not in the text. Writing “use Astra” changes nothing. Pick Sol by default, Luna for bulk tasks, Astra for the hardest ones.

Effort instead of “think harder”

Instead of “think very carefully” raise the reasoning level (High / Extra High). It works more reliably than adjectives in the prompt.

Ask for evidence

GPT-6 makes fewer factual mistakes, but not zero. Ask it to mark facts with sources and label estimates - it makes review faster.

Lesson 13 covers model choice and reasoning levels.

Check your understanding

A report has the right percentage formula but totals PLN 6500 instead of PLN 3400. What do you check first?

This is a self-assessment. Mark the lesson complete after doing the exercise and reviewing the result.

Do this now

Improve one vague prompt: describe the output, quality criterion, limits, sources and stopping condition. Compare both answers and record what actually improved.

How to know you can do it

Keep as evidence: Two answers compared on the same data.

  • Sources and rules are explicit.
  • You judge the result, not prompt length.
  • You can spot a duplicate or unsupported conclusion.

Mark the lesson complete after the exercise and review. Reading progress and quiz answers alone do not confirm practical skill.

Lesson status
In progress — 0% of the page scrolled
Next