The File That Is a Game Changer for Claude
What you’ll learn
Why a context file is not the same as a prompt, and why that distinction matters
The three jobs every context file needs to do (most people only do one)
Why legal requires a fundamentally different setup than every other function
The file that tells Claude everything it forgets between sessions
If you want AI to work inside a real legal function, you need to stop treating it like a search engine with chat.
You need to treat it like a new colleague on day one. Except this colleague forgets a lot of things when the conversation ends.
That's the problem context file (aka CLAUDE.md) solves.
I built mine this morning. It started with a conversation with our VP of Finance, who had been doing the same thing. His version was tight. Clear job description, output preferences, a formatting rules and continual updates. It worked well for finance.
I looked at his and immediately realized mine needed to be different. Finance has math. Math has right answers. You can verify an output against a known total. Legal does not work that way. The same term means different things in different jurisdictions. The same risk looks different depending on the company's tolerance for it. The same question has five defensible answers depending on which way you want to argue it.
General AI does not know any of that about your work. Without a context file, it defaults to the most common interpretation of every term and the most generic version of every answer. That is fine for a lot of use cases. It is not fine for legal.
What context file actually is
CLAUDE.md aka the context file is a plain text file that Claude reads at the start of every working session. It replaces the context you would otherwise re-establish through repeated prompting. It is not a prompt. It is a context document.
The difference matters. A prompt tells Claude what to do on a specific task. A context file tells Claude who you are, what you cover, how you think about risk, what good output looks like in your context, and what you do not want. That context changes every output, not just the one you are actively working on.
The three things it needs to do
Everything in a good context file is doing one of three jobs.
Guardrails
What Claude should never do without checking first. What sources it cannot use. What terms it needs to flag. These are the non-negotiables that prevent expensive mistakes.
How you want it to think
This is the section most people skip, and it is the most important one for legal. It covers your company's risk appetite, your jurisdictional flags, your preferred framing. It tells Claude when to be conservative and when to just advise on options. This is where you encode the judgment layer that makes legal advice actually useful.
The specifics
What you cover. What you do not cover. What your outputs need to look like. Source citation standards. Confidence scores. Risk ratings.
Without all three, you get a tool that understands tasks but not context. The outputs will be competent and generic. That is not what you are building.
The part that surprised me
The jurisdictional flags section does more work than I expected.
Example: Worker is a distinct legal category in UK employment law, sitting between Employee and Independent Contractor. It carries specific statutory rights. In US law, worker is used casually to mean anyone who works. Same word. Completely different legal meaning.
General AI defaults to the casual US interpretation unless you tell it not to. I built the flag into my file because I cannot afford to have it missed, and I should not have to catch it every time. I give it the context that lives in my brain and is critical to solving problems.
That is the point of the whole file. You are not just telling Claude what to do. You are telling it what you already know, so it works like a trusted colleague.
The practical takeaway
You do not need to be technical to build one of these. You need to be specific.
Start with three questions.
Who am I?
What do I actually work on?
What does a good output look like in my context?
Then add the edge cases. The jurisdictional flags. The format preferences. The things you keep correcting. Then ask Claude to help you build the file. Have Claude ask you questions that you can answer and include in the file.
I also suggest also adding these three sections:
Planning
Enter plan mode for any non-trivial task (3+ steps or architectural decisions)
Write detailed specs upfront to reduce ambiguity
If something goes sideways, stop and re-plan immediately — don't keep pushing
Use plan mode for verification steps, not just building
Subagents
Use subagents to keep the main context window clean
Offload research, exploration, and parallel analysis to subagents
One task per subagent for focused execution
Core Principles
Simplicity first: Make every change as simple as possible; touch minimal code
No laziness: Find root causes — no temporary fixes
Minimal impact: Only change what's necessary; avoid introducing bugs
Verify before done: Never mark a task complete without proving it works
Demand elegance: For non-trivial changes, ask "is there a more elegant way?" — skip for simple obvious fixes
Autonomous bug fixing: When given a bug report, fix it — point at logs/errors/tests and resolve them
That document is the foundation of your AI-native legal function.
If you want the full section-by-section breakdown and a template built from my file, it's in the Field Kit inside the Field Guide.
