# Writing the AI prompt for a filter

Source: Tender Navigator Knowledge Base, https://tendernavigator.com/knowledge-base/writing-the-ai-prompt (canonical HTML). Category: Filters. Published 2026-09-08, last updated 2026-09-08.

> The AI prompt on a Tender Navigator filter is a free-text description of what makes a tender a good fit for your company. Every tender the filter picks up is evaluated against this text, so the prompt should say what you sell, what a good fit looks like, what rules a tender out, and where you actually win.

## What the prompt is for

Keywords, countries and codes decide which tenders a filter reads. The **AI prompt** decides what the AI says about each of them. The model receives the tender notice and your prompt, and returns a verdict (Matched or Unmatched) with a reasoning that refers back to your criteria. The prompt is therefore the single most important text in Tender Navigator: a vague prompt produces vague verdicts, and a precise one produces reasoning you can act on.

The filter editor labels it **AI prompt** with the guidance "Describe what makes a tender relevant to you."

## What to include

Write it as you would brief a new colleague who has to pre-screen tenders for you.

1. **What you sell, concretely.** Services, products, the technologies or methods you use. "Managed cloud services on Microsoft Azure for organisations with 200 to 5,000 staff" beats "IT services".
2. **What a good fit looks like.** The buyers, project sizes, contract types and regions where you win. Name the signals in a notice that mean "this is for us".
3. **What rules a tender out.** Hard exclusions: sectors you do not serve, mandatory certifications you lack, minimum turnover thresholds you do not meet, on-site requirements you cannot fulfil, languages you do not work in.
4. **Where you actually win.** If your track record is in a niche, say so. The model can then prefer that niche over generic matches.
5. **How strict to be.** If you would rather see borderline tenders than miss one, say "when in doubt, mark as matched and explain the doubt". If your team is short on time, say the opposite.

## What to leave out

- Marketing language and adjectives. "Innovative", "leading" and "best-in-class" carry no evaluation criteria.
- Keywords lists. The keywords field already selects the tenders; repeating them in the prompt adds nothing.
- Instructions about formatting. The verdict and reasoning have a fixed shape.

## A structure that works

- **We are:** two or three sentences on the company and what it delivers.
- **A good fit:** a short list of positive signals.
- **Not a fit:** a short list of exclusions.
- **Preferences:** what to prioritise when several fits compete, and how to treat uncertainty.

## The seeded prompt and its TODO line

A filter created with **Save as Filter** or **Generate filter** arrives with a seeded prompt. From a search, the prompt restates the search and ends with a line beginning with **TODO** that asks for your business context; from the Business Profile, the prompt is assembled from your profile's regions, offerings, target buyers, capacity and evaluation notes. Read it, replace the TODO line with your own words, and remove anything that is not true for you before switching the filter on.

## Sharpening a prompt

When a verdict looks wrong, the reasoning on the result tells you which of your criteria the model applied. Adjust the sentence it relied on rather than adding new ones:

- Too many Matched results that you would never bid on: add exclusions, and name the disqualifying signals the reasoning missed.
- Good tenders marked Unmatched: check whether the prompt states a requirement more strictly than you mean it, and say how to treat uncertainty.
- Reasoning that repeats generic phrases: the prompt is probably generic too; add concrete examples of work you have delivered.

Changing the prompt affects tenders evaluated from the next daily run onwards. Existing results keep the reasoning they were given and show the filter snapshot they were evaluated against.

## Where the assistant can help

The [AI assistant](/knowledge-base/ai-assistant) in the dashboard can help you write a stronger filter prompt and read the reasoning on a result. It suggests text; you decide what goes into the filter.