Prompting techniques

Once a request is clearly specified, these are the techniques that change what you get back. Each has a job; using all of them at once mostly produces long prompts and mediocre answers.

Give it a role

“You are reviewing this as a procurement lead” narrows vocabulary, priorities and what counts as important. It works because it changes what the model treats as relevant, not because it grants expertise.

Worth using when the same document could be read four ways. Not worth using when the task is unambiguous — “translate this to Nepali” does not need a persona.

In Dhanu, a role you use repeatedly belongs on an AI staff member rather than at the top of every message.

Show examples

The most reliable technique there is. Give one to three examples of input and the output you want:

Turn each note into a task line.

Note: "call Sita re: invoice 2213"
Task: Call Sita about invoice 2213 — owner: me — due: today

Note: "ask legal if the NDA covers subcontractors"
Task: Ask legal whether the NDA covers subcontractors — owner: me — due: this week

Note: "renew domain before it lapses"
Task:

Examples pin down format, tone and edge cases simultaneously. If your examples disagree with your instructions, the examples win — so keep them honest.

Ask for the reasoning, when it earns its place

For anything with steps — a calculation, a comparison, a diagnosis — asking the model to work through it before answering measurably improves accuracy.

Work through the numbers step by step, then give the final figure.

Do not use it for simple retrieval or rewriting; you get a slower answer wrapped in narration you did not want. And if you want the thinking but not in the final text, say “reason it through, then give me only the conclusion”.

Break the work up

A request that contains “and then” three times is three requests. Long chains fail at their weakest link, and you cannot tell which link broke.

Ask for the outline, check it, then ask for the draft. In Dhanu you can also let the assistant plan and delegate the pieces itself — see Plans, subtasks and delegation — which is the same idea with the bookkeeping handled for you.

Ask it to check its own work

Now review that draft against the source document and list anything you cannot support with a quote.

A separate critique pass catches things the writing pass missed, because generating and evaluating are different jobs. It is most valuable on anything numeric, legal, or about to be sent to a customer.

Ask for structure

If the answer is going to be processed — pasted into a sheet, turned into tickets, compared with another answer — specify the structure exactly: columns, field names, one item per line. Vague structure is where “it almost worked” comes from.

For real structure, ask for the artifact itself. “Put this in a spreadsheet” gives you a live grid with formulas, which beats a markdown table that someone then has to retype.

Let it ask you

If anything is ambiguous, ask me before you start.

Cheap, and it surfaces the assumption that would otherwise have cost you a full rewrite. Particularly useful for a first attempt at a recurring piece of work that you plan to save as a prompt afterwards.

Choosing between them

Situation	Reach for
Output format keeps coming back wrong	Examples
Answer is confidently wrong on numbers	Step-by-step, then a self-check
Task is long and keeps drifting	Decomposition
Same request every week	A saved prompt, or an AI staff member
Tone is off	One example of the tone you want
You are not sure it understood	Let it ask you first

Next

Tips and tricks — smaller habits that compound.