Guides
Advanced principles
How to think about prompts you keep — putting durable context where it belongs, writing instructions that survive being reused, grounding rather than trusting, and testing a prompt like a small piece of software.
Everything so far is about one request. This page is about the prompts you use again — the staff instructions, the saved prompts, the standing instructions on a project. They are closer to software than to messages, and they reward being treated that way.
Put context where it belongs, not in every message
Dhanu has four places for durable context, and choosing the right one is the whole skill.
| Where | For | Applies to |
|---|---|---|
| Brain | Facts about your organisation | Everything that searches it |
| AI staff instructions | How one role should behave | Every run of that staff member |
| Project description and instructions | The situation this work sits in | Every task in the project |
| Saved prompt | A request you make repeatedly | Wherever you invoke it |
A fact that keeps getting retyped into messages is a fact that belongs in the Brain. An instruction you keep repeating to the same staff member belongs in its instructions. Prompts get shorter as the durable parts move out of them.
Ground it, do not trust it
The strongest thing you can do to a prompt is attach the source. The second strongest is to tell it what to do when the source is silent:
Answer only from the attached documents and the Brain. If they do not cover something, say what is missing instead of filling it in.
This turns the most dangerous failure — a fluent invention — into the most useful output: a gap you can go and fix. Dhanu’s own knowledge tools work this way, which is why an answer arrives with citations and a list of what was not found.
Write instructions that survive reuse
A prompt written for today quietly assumes today. Reusable ones avoid:
- Deixis — “the attached file”, “as we discussed”, “this quarter”. Say which file, which decision, which quarter, or leave a field for it.
- Buried defaults — if the prompt only works when the input is a PDF, say so in the prompt.
- One-off tone — “make it fun for the launch” ages badly in a template used for compliance updates.
Where a specific varies, make it a field: saved prompts support fill-in-the-blank
values, so {{client}} and {{quarter}} get asked for at the moment you use it.
Order matters more than length
Put instructions before the material they apply to, and put the constraint that must not be violated last as well as first. Long inputs get skimmed in the middle; the beginning and the end are where an instruction survives.
Separate the material from the instructions
When you paste a document into a message, mark it:
Instructions: summarise the policy below for new joiners.
--- POLICY ---
<the text>
--- END POLICY ---Without a boundary, a document that itself contains instructions (“ignore previous guidance and…”) is indistinguishable from your own. This matters most for material that came from outside your organisation — a supplier PDF, a scraped page, an email from a stranger.
Test a reusable prompt on its hard cases
Before you save a prompt or wire it into a staff member, run it three times:
- A typical case — does it do the obvious job?
- An empty case — a document with none of the information. Does it say so, or invent?
- An awkward case — the messy real example everyone remembers. Does it degrade sensibly?
The empty case is the one that matters. A prompt that fabricates when starved is a prompt that will eventually fabricate on something you send to a customer.
Fewer, sharper instructions
Long instruction lists dilute. Twenty rules mean the model is weighing twenty things, and the three that mattered get the same weight as “use British spelling”. When a prompt stops working, the fix is usually to delete half of it, not to add a paragraph.
Review what you have written down
Standing instructions rot: the policy changed, the client left, the format moved on. Whatever you keep — staff instructions, project instructions, saved prompts, Brain pages — is worth re-reading once a quarter. A confidently wrong standing instruction is worse than none, because nobody is looking at it.