The five context files: identity, voice, anti-slop, domain, stakeholder

Every file Claude can't find, it has to guess. Guesses cost tokens and produce generic output. The inverse is also true: every file you give Claude that it actually needs, you get back in output quality. Context files compound — the more you build, the less you re-prompt.

There are five files worth building in a specific order. Each one addresses a different failure mode.

The five files

1. Identity file

Who you are, your role, your current priorities, and the decisions already made. This stops Claude from making assumptions about your context that don't match reality. A consultant and a product manager at the same company should get fundamentally different outputs from the same prompt — the identity file is what makes that happen.

If you've already written your CLAUDE.md, you have the start of this file. The identity file can live inside CLAUDE.md or in context/identity.md, referenced from it. For long-running projects, a separate file is easier to update.

2. Voice file

How you write. Not "professional" — that's meaningless. Show Claude the actual texture of your writing with three to five examples of output you were happy with alongside three to five examples of writing that didn't sound like you. Side by side.

The contrast is what teaches. "Here's an email I sent that sounded right. Here's one that came out too formal. Here's the diff." Claude learns your voice from examples faster than from descriptions.

3. Anti-AI-slop file

This file earns its existence after your first Cowork session. You'll ask for a draft, and the output will contain em dashes, hedging phrases like "it's worth noting" or "at the end of the day," constructions like "it's not just X — it's Y," and possibly three exclamation points in the conclusion.

The anti-slop file is an explicit ban list. Em dashes: banned. Sentences that start with "Ultimately": banned. Bullet points that begin with "Leverage": banned. Whatever tells you hate, name them here, with examples of the wrong pattern and what to write instead.

The voice file and anti-slop file work together. The voice file shows Claude what you want. The anti-slop file bans what you don't. Both are needed — the voice file alone isn't enough because Claude has trained defaults that drift back toward generic output if you don't block them explicitly.

4. Domain file

Your industry terminology, the frameworks you use, your methodology, the jargon that has specific meaning in your context. If "pipeline" means something specific in your work, put that here. If you use a particular framework for competitive analysis or project planning, name it. If there's a term your clients use that doesn't match the industry standard, clarify it.

This file prevents the subtle wrongness that's hard to articulate — outputs that are technically correct but feel off because they're using the right words wrong, or the wrong words for right concepts.

5. Stakeholder file

The people Claude needs to know about. Your manager and what they care about. Your key clients and their communication preferences. Your team and who owns what. When you ask Claude to "draft an update for Marcus," it should know that Marcus is the CFO, wants numbers upfront, and doesn't read anything longer than one page.

This file also covers relationship context: "Emma and I disagree on the scope of this project — frame recommendations neutrally, don't take sides."

Where to store them

Your CLAUDE.md can include these files directly (for short files) or reference them (for longer ones). Referencing is better for anything over a few hundred words:

# Context Files
- Voice guidelines: see context/voice.md
- Banned patterns: see context/anti-slop.md
- Domain terminology: see context/domain.md
- Team and stakeholders: see context/stakeholders.md

Then in your task prompt: "Read CLAUDE.md and all files referenced in it before starting."

Build order

Start with identity and voice. These two files address the majority of output failures — wrong context, wrong register.

Build the anti-slop file after your first session. The patterns it needs to ban will be obvious from what Cowork produced.

Add domain and stakeholder files as you do real work. These files improve through use, not through up-front planning. Start sparse and add specifics as you notice Claude making wrong assumptions.