---

These four template files solve the problem. They give your AI a memory, a personality, and clear boundaries. Set up once, they work across sessions.

The 4 files at a glance

File	Purpose	Core question
IDENTITY.md	Role & personality	How should the AI act?
STYLE.md	Tone & formatting	How should the AI express itself?
CONTEXT.md	Project background	What does the AI need to know?
RULES.md	Hard limits	What must the AI never do?

IDENTITY.md — role, personality, working style

What goes in?

The Identity file defines who the AI is in your project. Not in the sense of role-play, but in the sense of: what stance should it take? How should it make decisions?

Core elements:

• Role: What's the job? (e.g. "Senior frontend developer", "content strategist", "product consultant")
• Core tasks: the 3–5 most important things the AI should do
• Personality traits: proactive vs. reserved? Opinionated vs. neutral? Creative vs. analytical?
• Decision principles: what to prioritise? (e.g. "speed over perfection", "revenue over features")
• Working style: when to ask, when to act independently?

Why this improves results

Without identity, the AI behaves like a generic assistant — polite but without stance. With identity you get consistent, context-aware answers that fit your project. Instead of "Here are 5 options" you get "I recommend option A because it fits your goal X."

Common mistakes

• Too vague: "Be helpful" says nothing. Better: "Always propose exactly ONE next step."
• Too many traits: 3–4 strong traits beat 10 generic ones.
• Contradictions: "Be proactive" and "Always ask first" clash. Be precise about when each behaviour applies.

STYLE.md — tone, language, formatting

What goes in?

The Style file defines how the AI communicates. Tone, language, format — everything that shapes how the output feels.

Core elements:

• Base tone: Casual-technical? Professional-warm? Sober-precise?
• Language rules: which language for what? "Du" or "Sie"? Jargon allowed?
• Formatting: markdown structure, length, code comments
• Dos: concrete examples, active language, clear structure
• Don'ts: filler phrases, buzzwords, unnecessary intros
• Good/bad example: shows the AI concretely what you want

Why this improves results

The AI defaults to a generic "assistant tone" — overly polite, with phrases like "Of course!" and "Great question!". The Style file kills that. The good/bad example is especially powerful: one concrete example beats 10 abstract rules.

Common mistakes

• Only defining don'ts, no dos: the AI needs to know what to do INSTEAD.
• No example: abstract style rules are interpretable. Concrete examples aren't.
• Forgetting language: if you work in multiple languages, define clearly when each applies.

CONTEXT.md — project profile, tech stack, audience

What goes in?

The Context file gives the AI the domain knowledge it needs to work within the right frame.

Core elements:

• Project profile: name, type, status, elevator pitch
• Tech stack: a table of technologies per layer
• Architecture: folder structure, key patterns
• Audience: primary, secondary, anti-audience
• Current state: checklist — what exists, what's missing
• Priorities: what takes precedence right now and why
• Glossary: project-specific terms and acronyms

Why this improves results

Without context the AI guesses. It proposes React when you're using Vue. It explains basics you already know. It ignores your architectural decisions. With context it works inside your project instead of alongside it.

The glossary is especially important on projects with their own jargon. If the AI knows "Binder" is a player role in your project and not a folder, you avoid misunderstandings.

Common mistakes

• Too much at once: only include what the AI needs regularly. Move deep dives into sub-files.
• Not kept current: stale context is worse than no context. Update priorities and status regularly.
• Forgetting paths: in code projects, the AI needs the folder structure or it creates files in the wrong place.

RULES.md — hard limits for code, communication, and scope

What goes in?

The Rules file defines what the AI must never do. These aren't preferences, they're hard limits.

Core elements:

• Absolutely forbidden: delete data, hardcode secrets, push to production, incur costs
• Code rules: no any types, no console.log, no unused imports
• Communication rules: don't hallucinate, don't hide things, no over-engineering
• Scope rules: no scope creep, no unrequested refactors
• Privacy: no personal data in logs, no credentials committed

Why this improves results

AI models are inherently eager to please — they want to help and tend to do more than asked. That leads to scope creep, needless refactors, and sometimes dangerous actions (e.g. changing production data). The Rules file sets clear guardrails.

Especially important: the golden rule at the end — "When in doubt, ask, don't act." That shifts the AI's default from "just do it" to "verify first".

Common mistakes

• Too restrictive: if everything's forbidden, the AI can't work. Focus on things with real damage potential.
• No explanation: "No any" is a rule. "No any because TypeScript strict mode is our first defence against bugs" is better obeyed.
• Not project-specific: generic rules like "write good code" do nothing. "No class components, only functional with hooks" is concrete.

How to set up the files

Step 1: copy the templates

Create a .claude/ folder in your project root (or wherever your AI looks for context files). Copy the four templates in.

Step 2: fill in the placeholders

Each file has [PLACEHOLDER]s and HTML comments with explanations. Go through them in order. Only fill in what you're sure of. The rest can be added later.

Recommended order:

1. CONTEXT.md first — that's the foundation (project, tech stack, glossary)
2. IDENTITY.md second — how should the AI act within this context?
3. STYLE.md third — how should it express itself?
4. RULES.md last — what limits do we need?

Step 3: iterate

The files get better the more you work with the AI. When it does something that bothers you: write a rule for it. When it does something well: document the pattern.

Maintenance tips

Weekly check:

• Are the priorities in CONTEXT.md still right?
• Is the project status current?

Per new project:

• Refill CONTEXT.md from scratch
• Adjust IDENTITY.md (different role?)
• STYLE.md and RULES.md are often reusable — just add project-specific bits

On problems:

• AI talks too much? → STYLE.md: add a length limit + don't
• AI does too much on its own? → RULES.md: tighten scope rule
• AI doesn't grasp context? → CONTEXT.md: extend glossary and architecture
• AI is too generic? → IDENTITY.md: sharpen decision principles

Using with other AIs (not just Claude)

The templates work across platforms, but with differences:

Claude (Anthropic)

• Reads .claude/CLAUDE.md automatically as project context
• Tip: merge all 4 files into a single CLAUDE.md or reference them via @ import
• Supports multi-level hierarchy: global → entity → project

ChatGPT / GPT-4 (OpenAI)

• No native filesystem reading. Context must be pasted as Custom Instructions or System Prompt
• Tip: IDENTITY + STYLE into "Custom Instructions" (persistent). CONTEXT + RULES as the first message in the chat.
• Mind the character limit: Custom Instructions cap around ~1,500 chars. Cut aggressively.

Cursor / Windsurf / other code editors

• They read .cursorrules, .windsurfrules, or similar files
• Tip: drop the content of the 4 templates into the editor's rules file. Format is usually Markdown.

Gemini (Google)

• System Instructions in API mode, "Gems" in the web UI
• Tip: IDENTITY + STYLE as System Instruction. CONTEXT as the first turn.

General notes for other AIs

• RULES.md is the most important across all models. Every AI needs clear limits.
• Good/bad examples (from STYLE.md) work universally — all models learn better from examples than from abstract rules.
• Shorter = better for models with small context windows. When in doubt: prioritise RULES + CONTEXT, trim STYLE + IDENTITY.
• Test: every model interprets instructions differently. After setup, hand it a test task and check the output fits.