Skip to main content
Book a demo
ProductsServicesMaps and data
“Placepoint is our main platform for property information. In practice, we only use Placepoint now.”
Petter CedellGroup Director of Property Investments, Selvaag Bolig
“We see a clear efficiency gain with Placepoint, as much of the information we need on ownership, ownership structures, zoning purposes and encumbrances is gathered in one place. Several of us in the organisation use the tool daily and get a lot out of it.”
Kristian AkslandPartner, CMS Kluge

Why Placepoint

The company

“We have 25 salespeople using Placepoint every day to generate leads and find tenants.”
Henning BjørnsgårdHead of Sales and Development, EIR
“We use Placepoint daily for credit work. Among other things, we can identify ownership concentration in business parks.”
Lars Kristian TveterSenior Relationship Manager, Danske Bank

How to get the most out of Placepoint AI

We have already tuned Placepoint AI itself: which tools Claude chooses, how lookups are scoped, and how much data comes back in each answer. You do not set any of this yourself. What you control is the context Claude works in and the way you ask. Both affect the quality of the answer and how much of your Claude usage quota it uses.

Keep conversations short​

Claude remembers everything said and retrieved in a conversation, and uses it as working memory when it answers. This memory is large, but it gets worse the fuller it becomes. In its documentation of the context window, Anthropic writes that accuracy and recall drop as the conversation grows, and calls this “context rot”. Its advice is to keep the context small and relevant: as little information as possible, but enough to produce the answer you want (Effective context engineering for AI agents). This is a property of Claude, not of Placepoint AI.

For you, this means two things. Start a new conversation when you switch property, client or topic, even if the old one still answers. And give Claude good context from the start, such as instructions and standing documents in a project, instead of building it up over a long conversation. If a long conversation becomes unclear or slow, ask Claude to summarise what has been settled, and paste the summary into a new conversation.

Turn on memory so Claude remembers you between conversations​

Placepoint AI remembers nothing itself. The client you use does the remembering, and that is where you control what is stored. In Claude this is called Memory: Claude stores who you are and what you work on, and can search earlier conversations when you ask about something you have discussed before. Each project has its own memory, and Cowork, which runs locally on your computer, does not use it.

Whether memory is on from the start depends on your plan:

  • Free, Pro and Max: memory is on from the start in Claude Desktop, on the web and in the mobile app.
  • Team and Enterprise: an Owner must first turn it on for the organisation. Then each user must turn it on under Settings, Memory, with the Generate memory from chats toggle. Until both are done, Claude remembers nothing from the previous conversation.

Under Settings, Memory you can also see everything Claude has stored, and correct or delete it. See Use Claude's chat search and memory to build on previous context.

If you use Placepoint AI from another client, the same principle applies, with different settings:

  • Microsoft 365 Copilot: memory is on from the start. You control it under Settings and more, Chat settings, Personalization, Saved memories. See Manage Copilot Memory in Microsoft Copilot.
  • ChatGPT: the settings are under Settings, Personalization, Memory, with separate toggles for saved memories and for using earlier conversations. See Memory in ChatGPT.

Give Claude instructions once, not in every prompt​

The simplest and most effective step is to tell Claude who you are, what you work on and what the deliverable should look like. Then you do not have to repeat it in every conversation, and the answers are closer to what you need on the first attempt. This applies to all use of Claude, not only Placepoint AI.

  • Your own profile: click your initials in the bottom left corner, select Settings and fill in Instructions for Claude. The instructions apply to all your conversations. Enter your role, company, field, the language you want answers in and the form the deliverable should take. See Understanding Claude's personalization features.
  • Projects: group related work in a project, for example one client, one acquisition or one portfolio, and put the instructions and standing documents there. Every conversation in the project inherits them. See What are projects? and How can I create and manage projects?.
  • The whole organisation: on Team and Enterprise plans, the Owner and Primary Owner can set instructions that apply to everyone in the organisation, under Organization settings, Organization and access. Where they conflict with a user's own instructions, the organisation's instructions take priority. See Set organization instructions.
  • ChatGPT: you add instructions under Settings, Personalization, Custom Instructions. See Ask ChatGPT to use Placepoint for a text that makes ChatGPT use Placepoint before searching the web.

Anthropic's own advice is in Prompt engineering best practices, and the technical reference it keeps up to date is Prompting best practices. Four of the recommendations matter most: tell Claude what to do instead of what not to do, give the reason for each requirement so the model understands the purpose, allow it to say that something is missing, and keep the instruction short. Longer instructions do not give better answers.

Example instructions for an organisation​

A law firm that uses Placepoint AI for factual groundwork, not for advice, could add something like this for the whole organisation. Anthropic recommends the tags because they keep the different requirements apart, and because you can later change one block without touching the rest.

<rolle>
You provide factual groundwork to lawyers. We make the legal assessment ourselves,
so the answer should be verifiable rather than conclusive.
</rolle>

<kilder>
State the register and date for every factual claim about a property, and keep
Grunnboken data separate from matrikkel data. Mark your own reasoning with
"Inference:", so we can see what the register says and what you derive.
If something is not in the register, write that it is not registered.
</kilder>

<nar_noe_mangler>
If you lack context that would materially change the answer, ask one precise question
before you answer. Reasonable inferences are fine when they are marked as such. Figures,
names, quotations and source references must come from data you have retrieved.
</nar_noe_mangler>

<personopplysninger>
You do not show personal names or national identity numbers: Placepoint AI removes them.
If we need an identified registered owner, we look it up in Placepoint Fusion,
where access requires a legitimate interest, or in Grunnboken.
</personopplysninger>

<leveranse>
Answer in plain text or a table that we paste into our own templates. Ask first
if the task looks like it needs a report, PDF or dashboard: it takes longer,
and we rarely use it. End with findings, open questions and what should be
investigated further.
</leveranse>

The instruction does three things. It says what Claude should do, not what it should avoid. It gives the reasons for the requirements. And it gives the model a way out when the evidence is thin, so it does not force an answer. These three techniques matter most.

Replace the content with what applies to your industry. A property company might ask for portfolio figures per SPV, for example, and a bank for a traffic-light rating per risk topic.

Standard instructions by industry​

The blocks below use the same three techniques for each industry, and are split into three parts: who you are and what Claude should deliver, what to prioritise, and what form the deliverable should take. Each example is written as if it came from one organisation in the industry. Select your industry, and replace whatever does not fit.

<rolle>
We are a regional bank with business customers, and you provide factual groundwork
for collateral and credit assessment of the properties pledged as security.
We make the credit decision ourselves, so the answer should be verifiable
rather than conclusive.
</rolle>

<prioriter>
Registered owner and form of ownership first, then encumbrances from Grunnboken:
mortgages with amount and priority, easements, restrictions on disposal and
prohibitions on subdivision. Then natural hazards for each property: flood,
landslide, quick clay and storm surge, with hazard awareness maps kept separate
from hazard zones. Finally tenants with lease terms where the rental income
services the loan. State the register and date for every item.
</prioriter>

<leveranse>
A table with one row per property and one column per topic: encumbrance, natural
hazard, tenant, source, date. Write "not covered by the register" where coverage
is missing, rather than leaving out the row. End with the three things we should
clarify with the customer before the case goes to the credit committee. No
recommendation on the loan or loan-to-value ratio.
</leveranse>

If several of you work in different ways, put the shared part on the organisation and the industry block in each project.

Send only what the lookup needs​

What Claude sends in a lookup reaches us and is logged, see What does Placepoint log when an AI tool retrieves data?. Placepoint AI usually only needs what identifies the property, company or area: address, matrikkelnummer (cadastral number), organisation number, place name, municipality or coordinates. The exception is lookups of roles in companies, where the name of the managing director, board member or shareholder is the lookup itself. National identity numbers, customer numbers, policy numbers, claim numbers and text from your own documents do not improve the answer, and should stay in the AI tool.

This matters most when Claude has access to your own documents, in a project or through another connector. The model may then include text from the document when it builds the lookup. Add a block like this to the instructions for the organisation or project:

<oppslag_mot_placepoint>
When you use Placepoint AI, send only what identifies the property,
company or area: address, matrikkelnummer, organisation number,
place name, municipality or coordinates. Only send personal names
when the lookup concerns roles in companies. Keep national identity numbers,
customer numbers, policy numbers, claim numbers and text from our own documents
out of the lookup, because Placepoint does not need them and everything sent
is logged there. If the document only gives you a personal name, ask me for
the address instead of searching on the name.
</oppslag_mot_placepoint>

The block uses the same techniques as the examples above: it says what Claude should send, gives the reason, and offers a way out when the document does not contain what the lookup needs. An instruction guides what the model sends, but it is not a guarantee.

Ask for what you actually need​

  • One task per conversation. Start a new conversation when you change topic. A long conversation carries everything retrieved earlier, and that costs both time and quota without improving the answer; see Keep conversations short.
  • Say what form you want. If you need three figures, ask for three figures. A report, table, map or dashboard costs more than a short answer. Reports are delivered as HTML by default because it is fastest; ask for PDF when you are actually going to send the document on.
  • Narrow it down. One matrikkelnummer, one street, one municipality or one organisation number gives faster and more precise answers than “Oslo”.
  • Reuse the recipe. When an analysis works, ask Claude to write the prompt down as a recipe you can run on the next property or portfolio. Put it in the project instructions.
  • Choose the model for the task. Use the most advanced model to build the structure and the recipe, and a faster model to run the same recipe many times.

Ask for sources and read the caveats​

  • Require a source and date for every factual claim. Claude will readily state which register the information comes from and when it was retrieved, and link to the map layer or the plan document. Make it a standing requirement, and the answer becomes verifiable.
  • Ask Claude to separate what the register says from what it derives itself. This is the most important distinction in factual groundwork.
  • Definitions vary between municipalities. “Ubebygd” (undeveloped) and “sentrumsformål” (town-centre use) do not mean the same in every municipality. Ask Claude to flag which definition it has used when the answer covers several municipalities.
  • If data is missing, the answer should say so. Placepoint AI tells you when there is no plan data for a property, instead of describing it as unregulated.
  • Claude can be wrong. The results are a starting point, not a final answer, and must be checked against official sources before you use them for decisions.

If an administrator has added the connector for the whole organisation, each user must still sign in with their own Placepoint user account the first time. Access follows the user, not the URL. How to connect is described on Placepoint AI.

See also​

Sist oppdatert
These pages are generated automatically and may contain errors.This applies to screenshots and videos in particular. See About these help pages. Let us know through "Did you find what you were looking for?" at the bottom of the page, the chat at bottom right or support@placepoint.no.