How to give Claude context about your work

Every morning you open a new conversation and the model knows nothing. Not what you shipped yesterday, not the argument in the channel about the schema, not the decision your lead made on the call you missed. So you spend the first five minutes reconstructing it by hand.

That reconstruction is a tax, and it compounds: the context you paste is only ever the context you remembered to look for, which is a fraction of what exists and biased toward whatever you happened to read last. Here are four ways to reduce it, from cheapest to most complete.

1. Write the context down once, in the repo

The lowest-effort improvement is a file the assistant reads automatically. Claude Code picks up CLAUDE.md; other tools have their own convention. Put in it the things you would otherwise explain every time: what this project is, the conventions that are not obvious from the code, the decisions someone would otherwise undo.

This works well for facts that are stable and about the code. It works badly for anything that changes weekly or lives in a conversation — which is most of what you actually need.

2. Paste better, not more

When you do paste, paste the argument rather than the conclusion. A model given the decision alone will defend it; a model given the thread that produced it can tell you the tradeoff was already considered and rejected for a reason that no longer holds.

• Include who said it. Attribution changes how much weight a claim deserves, for a model as much as for you.
• Include the date. Half of what is in a channel is superseded, and nothing in the text says so.
• Include the thing that was rejected, not just the thing that was chosen.

3. Point the assistant at a search tool

Better than pasting is letting the model fetch. If your assistant supports tools, giving it a search function over your own sources removes the step where you have to know in advance what is relevant.

The catch is that you now own a retrieval pipeline: a connector per source, the auth for each, backfill, incremental sync, and an index that goes quietly stale when a token expires. This is a real project, and it is why most people who start here end up back at pasting.

4. Use a context server

The Model Context Protocol standardised this. Instead of building a bespoke tool per assistant, you run one server that holds your context and speaks a protocol any MCP client can call. Claude Code, Claude Desktop, and your own agents all connect the same way.

What makes it work is coverage. A server over your documents alone gets you what you already had. A server over the conversations — Slack threads, email, what was said on the call — gets you the part that was never written down anywhere else.

❯ what did I miss on the payments bug?

jarvis － search_context(query: "payments bug")
  └ 12 messages across 4 sources

Tom's fix is in review, not shipped. It double-charged 3 customers.
  #eng · 3d   ENG-2481   PR #142   Eng sync, Aug 19

Note the citations. A context server that answers without telling you where the answer came from has replaced one trust problem with a worse one — you can no longer check it. Insist on sources.

What to do about it

Start with the file in the repo, because it costs an hour and removes the most repetitive explaining. Then look honestly at what is left: if what you keep pasting is threads and call notes rather than facts about the code, no amount of documentation will fix it, because the problem is retrieval, not writing.