A good brief. A verified result.

The shortest method is: define the result, provide context, set boundaries, verify execution. Below is a practical approach to writing, planning, analysis and building websites.

1. Describe the result, not just the topic

“Help me with marketing” does not say what should be produced. A better assignment specifies the audience, format and goal: “Prepare three versions of a short service description for a small-business owner. Each should be no more than 80 words, avoid unsupported promises and end with a specific next step.”

Before writing a prompt, answer three questions:

• What should be produced?
• Who will use it?
• How will I know the result is ready?

Clear criteria reduce rounds of corrections in which AI changes things nobody requested.

2. Provide context that changes the answer

Provide the starting state, relevant facts, constraints and an example of the tone or format. If you refer to a particular document or page, give its exact link or name. Do not make the model guess which of five similar files you mean.

You do not need to paste the entire history. Include what affects the decision. Too much material makes important information harder to find, and unnecessary data increases risk. Keep secrets, passwords and private details out of the prompt.

3. Set the format, constraints and boundaries

Say whether you expect a list, table, finished text, code or a short recommendation. List important constraints: length, budget, permitted sources, language, elements to preserve and things the AI must not change.

AI can draft a message, publication plan or page. Sending a message, publishing, buying or changing real data is a separate step: first check the content, recipient and scope of authorisation.

4. Divide large tasks into stages

For larger work, establish a sequence: investigation, draft, execution and verification. Each stage should have an owner, a specific input and acceptance criteria. One reliable place for the case status helps prevent parallel versions and repeated searches.

Delegate only the part the other person or tool can perform correctly. Receive and check the result before moving on. Adding more links to a handoff chain does not in itself make it a better workflow.

In the mode marked by the command LP the more capable model prepares only a ready-to-use prompt for Luna. It does not send it, start the task, supervise it or check its result; the user passes on the prompt and receives the result. Any cost benefit remains hypothetical until prompt preparation, Luna's work and possible corrections are included and measured.

5. Require uncertainty to be named

Ask for confirmed facts, inferences and unknowns to be separated. When a subject changes quickly, check current primary sources and request links supporting important claims. A confident tone is not a source, however much it may sound like one.

If sources conflict or something cannot be checked, preserve that uncertainty in the result. Do not replace “I don't know” with a number, date or promise invented to make the answer look complete.

6. Check what works, not just what is claimed

Match the check to the task. For text, verify facts, links and compliance with the brief. For a spreadsheet, read back saved cells and formulas. For a website, check desktop, tablet and phone layouts, clicks, refresh and the key function.

“File saved” does not always mean “change works”. Likewise, a running process does not prove that a result reached its recipient or that a public page was actually updated. Gather evidence appropriate to the goal.

7. Give specific corrections

Instead of “make it better”, identify the exact problem and desired effect: “The heading is too long. Shorten it to eight words, preserve the meaning and do not change the section below.” A good correction limits scope and protects what already works.

If the result will be reused, record the verified format, preference or procedure. Memory should shorten the path to the next result, not turn every prompt into a manual for the entire world.

8. Keep a person involved in decisions

AI is useful for preparing options and comparing arguments. Responsibility for significant decisions, risks and consequences remains with people. Appropriate oversight depends on the consequences: a text draft requires different checks from a bank transfer or public publication.

For important matters, ask the model to list risks, assumptions and missing information before recommending action. This often saves more than another round of correcting the consequences of a hasty decision.

9. Establish simple working rules

• Briefly first, details later. Put the most important conclusion first; add a plan and detail when they help a decision.
• One owner and a clear status. For larger work, establish who owns the result, what has been confirmed and the next step.
• Separate a draft from action. Preparing content does not mean sending it, and a mock-up is not a deployed website. Before an external action, check scope and required approval.
• Say what is fact. Separate sources, inferences and unknowns; mark missing data instead of filling gaps with guesses.
• Finish with evidence. After execution, check the relevant result and briefly state what was done and what remains.

Ready-to-use prompt template

Fill in only the fields that actually matter:

You do not need a longer prompt. It needs to be precise enough that the model does not have to invent the missing task.

See the full comparison and 9 graphics ↗

ORG-STRUCT-23 — Organising Work with AI
Three levels of documentation, one owner per case and a verified result: a simple way for AI to continue the right task without confusion.

When AI creates its own bottleneck
How the registry grew to 336.7 MB, what we simplified and what before-and-after measurements actually show.