How to use Claude Projects to organize a long research workflow

Claude Projects look like folders for chats, but the real power is in the persistent context: a project-scoped system prompt, a knowledge base you upload once, and curated conversation starters. This guide shows how I run a weekly competitive research workflow inside a single Project — and the limits that wasted the most of my time.

What you'll learn

• What Claude Projects actually persist between chats (and what they don't)
• The setup checklist: project instructions, knowledge upload, conversation starters
• A real workflow: weekly competitive research, end-to-end
• The two gotchas that quietly eat your context window

What Claude Projects actually do

A Project is a container that bundles four things:

1. A system prompt (called "Project instructions") that is prepended to every new chat inside the project. It is not the same as the global "Custom preferences" in your account settings.
2. A knowledge base of files you upload. They are available to any chat in the project, but Claude decides when to read them.
3. A library of past chats that took place inside the project, so you can scroll back to a thread from three weeks ago.
4. Conversation starters that appear as the first clickable options when you open a new chat.

What Projects are NOT: they are not a version-control system, they are not a way to share chats with teammates unless you are on a Team or Enterprise plan, and they are not a replacement for an external RAG pipeline. The knowledge base has a hard size cap (around 500k tokens at the time of writing), and uploaded files do not auto-update — if your source document changes, you must re-upload.

Setup walkthrough

1. Create the project

• In the Claude sidebar, click Projects → New Project.
• Name it. I name mine by use case, not by date: Competitive research — fintech, not Research week 23. The name shows up in every chat title and you will thank yourself later.

2. Write project instructions

This is the single most important step. Project instructions are a system prompt — they steer Claude's voice, scope, and output format for every chat. Bad instructions are vague ("be a research assistant"). Good instructions are concrete.

A template I use for a research project:

You are helping me with weekly competitive research in the [INDUSTRY] space.
For every response, prefer specificity over breadth. Cite sources when you
make a factual claim. When you don't know, say so explicitly. Use bullet
lists for comparison, prose for narrative. Default to English. If I ask
about a company, structure the answer as: (1) what they ship, (2) who
they sell to, (3) what changed in the last 30 days.

Paste that into Project instructions. It is editable at any time, but changes only apply to new chats — existing chats keep the instructions they were opened with.

3. Upload the knowledge base

The knowledge base is the second superpower. For my research project, I upload:

• A 1-page glossary of internal terms and acronyms (saves Claude from inventing them)
• The latest 4 quarterly earnings transcripts for each company I track
• A CSV export of my CRM's "opportunities lost to competitor" field

Claude does not always read these files — it decides based on the question. If you ask a question that needs a specific file, name the file in your prompt ("check the earnings transcript before answering").

4. Set conversation starters

Conversation starters are clickable prompts that show up at the top of a new chat. I use them to enforce a workflow. My current set:

• "Summarize what changed for [COMPANY] in the last 7 days."
• "Compare [COMPANY A] and [COMPANY B] on pricing and packaging."
• "Draft a Monday-morning brief for the team on a specific theme."

Starters are not magic — they are just text. But they save you from re-typing the same prompt in every chat.

A real workflow: weekly competitive research

Every Monday morning, I open my Competitive research — fintech Project and run this exact sequence:

1. Open a new chat. Use the first starter: "Summarize what changed for [COMPANY] in the last 7 days." Run it for each company I track (usually 6-8).
2. Open a second chat. Use the second starter to compare two companies on a specific axis. Use the comparison to update my internal "battle card" document (which I keep in Google Docs, not inside the Project — see gotcha below).
3. Open a third chat. Use the third starter to draft the Monday brief. I copy the output into Slack.
4. At end of day, I export the three chats as a single PDF (see the ChatGPT export tutorial — the same browser-print trick works in Claude) and attach it to the team's weekly digest email.

This takes about 90 minutes instead of the 4-5 hours it used to take when I was re-typing prompts and pasting context from scratch.

Gotchas

1. Knowledge base version staleness

If you upload a file in week 1 and forget about it, you will get confident answers based on stale data. Set a calendar reminder to re-upload key files monthly. There is no auto-sync.

2. Project instructions don't carry over to existing chats

If you change your project instructions mid-week, chats you opened yesterday still use the old instructions. This is a feature for reproducibility, but it surprises people. Open a new chat when you change instructions.

3. The context window is shared, not per-chat

The project-wide context window includes your instructions, all uploaded knowledge files, and the current chat. If you upload a 200k-token knowledge base, every chat starts with 200k fewer tokens available. The 500k cap is across the project, not per chat.

4. No good way to share Projects externally

You can share with teammates only on Team and Enterprise plans. There is no "view-only link" to send a Project to a non-Claude user. For external sharing, export to PDF and email.

FAQ

How much knowledge can I upload?

Roughly 500k tokens across all files in a Project. After upload, you see the actual token count. A 100-page PDF is typically 30-50k tokens.

Do Projects share data with my main Claude chats?

No. Knowledge and instructions are scoped to the Project. A chat in another Project, or in the main Claude UI, sees nothing from this Project.

Can team members see my Projects?

Only on Team and Enterprise plans, and only for Projects that are explicitly shared. Personal Pro accounts are private.

Why does Claude sometimes ignore my uploaded knowledge?

It decides when to consult uploaded files based on the question. If a file is relevant, it usually says "based on [filename]" in its answer. If you want to force a check, say so explicitly: "Open the Q3 transcript file and quote the relevant paragraph."

Can I export an entire Project, chats and all?

Not in one click. You can export each chat individually via the share menu, or use the browser-print trick in bulk (open each chat in a tab, print all to PDF).