Guide · Giving AI your background

A project takes your outputs from zero to 80%.

For anyone still opening a fresh chat and typing from zero every time.

A project is a folder inside Claude or ChatGPT that holds your organisation's key documents, like your strategy, reports and style guide, so the AI can reach for them when it needs them. It turns a chat window into a contextually aware ecosystem. Ask for a fundraising email in a bare chat and you get a template anyone could have written. Ask inside a project and you get something 80% there.

The idea

Picture an intern on their first day: bright, motivated, has read every book and every website, and knows nothing about your organisation beyond what is public. Send them off to work without an induction and the work will be poor, and it will not be their fault. A project is the induction.

The prompt is the how: the instructions. The context is the what: the material to work with. Most people polish the how and starve the what, and the what matters at least as much.

One mechanical fact shapes how you build one. The instructions field is read on every single message, which is why guardrails belong there. An uploaded file is not: when a question arrives it pulls three to five snippets out of your files, the ones it judges relevant. Instructions always on, files on demand.

The same prompt, twice. Once in a blank chat. Once inside a project holding your strategy, your voice and last year's report. Same words in, and nothing like the same thing out.

Copy the prompt
Draft a fundraising email for our current appeal.

You do not have to write the documents. The fastest way to fill a project is to be interviewed. It asks one question at a time, then writes your context files for you.

How it works

  1. Create the containerThink of it as a pyramid: one organisation-wide project underneath, then one per department, program or funder above it. The honest catch is that projects do not nest, so the context the top ones need has to be copied down by hand. Switch memory to project-only where offered, at creation: what happens in this container stays in this container, and the setting cannot be changed later.
  2. Write the instructions for the environmentKeep them generic to the space, not task prompts: who you are, how to speak, what never to invent. Task prompts live in a prompt library or a skill, called when needed. Three things make an instruction stick: wrap it in double asterisks so it reads as bold, use capitals for the one rule that matters most, and give the replacement rather than only the ban. Do not use em dashes, use commas instead.
  3. Add the onboarding documentsWhatever you would hand a new staff member: strategic plan, impact report, communication guidelines. Two are enough to start, the strategic plan and the communication guidelines. Curated: the essence, not the whole shared drive.
  4. Or get interviewed insteadThe fastest way to fill the context is not writing documents: it is having the AI interview you, one question at a time, and write them for you.
  5. Watch for clutterTwo names for the same failure: context clutter, context slop. Five years of history dumped in is worse than nothing: the model drowns. Give it enough to do the job well, and no more.

The sentences to type

Run the first one twice. It is the whole argument in two minutes.

The 80% test

Run this once in a temporary chat with no memory, and once inside a project holding your documents. Put the two outputs side by side. That difference is what context is worth.

Draft a fundraising email for our current appeal.
Write my instructions

Do not write the instructions field by hand. Describe the project out loud and have the AI draft them, then cut it back. Too much and trimmed beats blank.

I have set up a project for [organisation] and I want you to write its instructions. Here is what the project is for: [what work happens in here]. Here is who we are: [paste your About page]. The knowledge files in it will be [list them].

Write the instructions in raw markdown, using ## headings, with two sections: context about us that should be repeated into every conversation, and output format rules. Keep it under 8000 characters. Do not put task prompts in it.
Interview onboarding

The easiest way to fill a project: the AI writes the documents, you just answer.

Interview me to build this project's context. Ask one question at a time (who we are, what we do, how we speak, what we are working on) until you have enough to write an about-us file, a voice file and a current-projects file. Then write them, and do not invent anything: where you are not sure, ask.
The curation pass

For the forty-page report you were about to upload whole. Extract the essence instead.

Read this document and extract only what an AI assistant working for us would actually need: our goals, our constraints, our voice, the facts that recur. One page, plain language, written to be read by an AI as working context.