"CLAUDE.md Authoring Guide 2026 — What to Include, What to Avoid, and Five Anti-Patterns"

Four locations, loading rules, the 200-line ceiling, @import syntax, and when to split into .claude/rules/

핵심 요약

  • Audience: You've run /init and have a CLAUDE.md, but you're not sure what else to put in it.
  • What you'll get: Per the official /memory docs, the four locations, load rules, what goes in vs out, three effective-writing rules, @import, .claude/rules/ path-scoped splitting, and five anti-patterns to avoid.
  • Prerequisite: Claude Code installed (install guide), slash commands basics (reference).

1. What CLAUDE.md is

Claude Code has two memory systems that carry knowledge across sessions.

CLAUDE.md Auto Memory
Who writes it You Claude itself
What it contains Instructions, rules Learnings, patterns
Scope Project / user / org Per working tree
Loaded Every session start Every session start (first 200 lines or 25KB)
Best for Coding standards, workflows, architecture Build commands, debug insights, preferences

CLAUDE.md is what you write to guide Claude's behavior. Auto memory is what Claude writes based on your corrections. Both are injected as context, not system prompts — advisory, not enforced. The more specific and concise, the more reliably Claude follows them.

This post focuses on CLAUDE.md.

2. Four places CLAUDE.md can live

Scope Location Purpose Shared with
Managed policy macOS: /Library/Application Support/ClaudeCode/CLAUDE.md
Linux/WSL: /etc/claude-code/CLAUDE.md
Windows: C:\Program Files\ClaudeCode\CLAUDE.md		 Org-wide policy (deployed by IT/DevOps) All users in the org
Project ./CLAUDE.md or ./.claude/CLAUDE.md Team-shared project instructions Team via version control
User ~/.claude/CLAUDE.md Personal preferences across all projects You only
Local ./CLAUDE.local.md Per-project personal notes (gitignore required) You only (this project)

Priority: more specific wins — project > user > managed. But note that content is concatenated, not overridden. Within a directory, CLAUDE.local.md is appended after CLAUDE.md, so on conflict, your personal notes come last (and win).

3. Loading rules — how the directory tree is walked

Claude Code walks from your current working directory up the tree, reading CLAUDE.md and CLAUDE.local.md at every level. Example:

/home/me/monorepo/          ← CLAUDE.md (shared)
├── foo/                    ← CLAUDE.md (foo team)
│   └── bar/                ← CLAUDE.md (bar feature)
│       (working directory)

Running Claude here loads all three files concatenated into context.

3.1 Subdirectories are on-demand

Conversely, CLAUDE.md in sub-directories is NOT loaded at launch. It's loaded when Claude actually reads a file in that subdirectory. This keeps unrelated sub-project instructions out of context in monorepos.

3.2 --add-dir does not auto-load CLAUDE.md

--add-dir grants file access but does not load the directory's CLAUDE.md by default. To opt in:

CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1 claude --add-dir ../shared-config

3.3 Excluding other teams' CLAUDE.md in a monorepo

If an ancestor CLAUDE.md is irrelevant to you, exclude via .claude/settings.local.json:

{
  "claudeMdExcludes": [
    "**/monorepo/CLAUDE.md",
    "/home/user/monorepo/other-team/.claude/rules/**"
  ]
}

Note: managed policy CLAUDE.md cannot be excluded (by design — organizational policy must always apply).

4. What to include, what to exclude

The official guidance is blunt. CLAUDE.md consumes tokens every session, so noise hurts adherence.

✅ Include ❌ Exclude
Bash commands Claude can't guess (e.g. pnpm dev, make seed-db) Anything Claude can figure out by reading code (e.g. scripts listed in package.json)
Code style rules that differ from language defaults Standard language conventions Claude already knows
Testing commands and preferred test runners Long API documentation (link to external docs)
Repository etiquette (branch naming, PR conventions) Info that changes frequently (versions, contacts)
Architectural decisions specific to the project Long explanations or tutorials
Dev environment quirks (required env vars) File-by-file codebase descriptions
Common gotchas and non-obvious behaviors Self-evident practices like "write clean code"

For every line, ask: "Would removing this cause Claude to make a mistake?" If not, cut it. A bloated CLAUDE.md buries the important rules.

5. Three rules for effective writing

Per the official best practices page.

5.1 Size — under 200 lines

Target under 200 lines per CLAUDE.md. Beyond that, context cost rises and adherence drops. The docs note: "If Claude keeps doing something you don't want despite a rule against it, the file is probably too long and the rule is getting lost."

This ceiling is for CLAUDE.md. Auto memory's MEMORY.md is a separate file — only its first 200 lines or 25KB are loaded, whichever comes first. Anything past that is ignored (Auto Memory splits content into topic files).

5.2 Structure — headers and bullets

- Use ES modules (import/export), not CommonJS (require)
- Destructure imports when possible

- Typecheck after a series of changes
- Prefer running single tests, not the whole suite

A scannable structure beats dense paragraphs.

5.3 Specificity — verifiable

Official examples:

Vague (❌) Specific (✅)
"Format code properly" "Use 2-space indentation"
"Test your changes" "Run npm test before committing"
"Keep files organized" "API handlers live in src/api/handlers/"

For critical rules, adding IMPORTANT: or YOU MUST improves adherence (official recommendation).

6. @import syntax

CLAUDE.md can import other files with @path/to/file.

See @README.md for project overview and @package.json for available npm commands.

- Git workflow: @docs/git-instructions.md
- Personal overrides: @~/.claude/my-project-instructions.md
  • Relative paths resolve from the file that contains the import (not the working directory).
  • Absolute paths allowed.
  • Recursive imports allowed, max 5 hops.
  • Imported files are still loaded at launch, so this is organization, not token savings. For actual context savings, use path-scoped rules in .claude/rules/.
  • First-time imports prompt for approval. If you decline, imports stay disabled and the dialog doesn't return.

7. Coexistence with AGENTS.md (e.g. Codex)

Claude Code does not read AGENTS.md directly. If your repo already uses AGENTS.md, import it from CLAUDE.md:

@AGENTS.md

## Claude Code

Use plan mode for changes under `src/billing/`.

Claude reads AGENTS.md first, then appends the Claude-specific section. One source of truth, no duplication between Claude Code and Codex.

8. .claude/rules/ — splitting for larger projects

When CLAUDE.md grows too large, move topic-specific rules to .claude/rules/.

your-project/
├── .claude/
│   ├── CLAUDE.md
│   └── rules/
│       ├── code-style.md
│       ├── testing.md
│       └── security.md

All .md files in .claude/rules/ are discovered recursively. Subdirectories (frontend/, backend/) are fine.

8.1 Path-scoped rules

Add a paths frontmatter field and the rule loads only when Claude reads a matching file. No context cost otherwise.

---
paths:
  - "src/api/**/*.ts"
---

- All endpoints must include input validation
- Use the standard error response format
- Include OpenAPI documentation comments

Multiple patterns + brace expansion:

paths:
  - "src/**/*.{ts,tsx}"
  - "lib/**/*.ts"
  - "tests/**/*.test.ts"

Rules without paths load unconditionally at the same priority as .claude/CLAUDE.md.

8.2 Share across projects with symlinks

ln -s ~/shared-claude-rules .claude/rules/shared
ln -s ~/company-standards/security.md .claude/rules/security.md

Symlinks are supported; circular links are auto-detected.

8.3 User-level rules

Put rules in ~/.claude/rules/ to apply them across all your projects. Priority: project > user.

9. Five anti-patterns

9.1 Bloated file

CLAUDE.md over 200 lines. Critical rules drown in verbiage and adherence falls off a cliff. - Fix: Prune. Ask "does removing this cause a mistake?" If not, delete. Review periodically to kill dead rules.

9.2 Contradictory rules

Bad example: "Use 2-space indentation" in one file + "Follow the Google style guide" (4-space) elsewhere. Claude picks one arbitrarily. - Fix: Periodically scan all CLAUDE.md files + nested ones + .claude/rules/ together to remove contradictions.

9.3 Generic hedging like "write clean code"

Self-evident platitudes dilute the specific rules. - Fix: Don't write what Claude already does by default. Ask why this rule needs to exist.

9.4 Procedures in CLAUDE.md

Multi-step workflows like "to deploy: 1) test 2) build 3) push…" eat tokens every session. - Fix: Move to a skill (/deploy). CLAUDE.md is for facts and rules, not procedures.

9.5 Assuming nested CLAUDE.md survives /compact

The project-root CLAUDE.md is re-injected automatically after /compact (officially documented). Nested CLAUDE.md in subdirectories is NOT re-injected — it reloads only when Claude reads a file there next. - Fix: If an instruction was only given in conversation, migrate it to CLAUDE.md. Don't design nested CLAUDE.md assuming post-compact availability.

10. Getting started — /init

/init

Claude analyzes your codebase (build system, test framework, code patterns) and produces a starter CLAUDE.md. If one already exists, /init suggests improvements rather than overwriting.

Interactive multi-phase flow via env var:

CLAUDE_CODE_NEW_INIT=1 claude

This mode asks which artifacts to set up (CLAUDE.md, skills, hooks), explores the codebase with a subagent, follows up for gaps, and presents a reviewable proposal before writing anything.

After generation, treat CLAUDE.md like code: review when things go wrong, prune regularly, test by observing whether Claude's behavior actually shifts.

11. Diagnostics — /memory and the InstructionsLoaded hook

11.1 /memory — see what got loaded

Running /memory in-session lists every CLAUDE.md, CLAUDE.local.md, and rules file loaded for the current session. If a file isn't in the list, Claude can't see it — first thing to check is location.

11.2 InstructionsLoaded hook — debug via event log

To trace when path-scoped rules or lazy-loaded subdirectory files actually load, use a hook:

{
  "hooks": {
    "InstructionsLoaded": [
      {"type": "command", "command": "echo $CLAUDE_INSTRUCTIONS_PATH >> /tmp/claude-rules.log"}
    ]
  }
}

Full hooks guide coming in the next post.

12. Counter-scenarios — when this guide doesn't fit

  • Prompt-only workflow → Skip CLAUDE.md. Pass instructions via --append-system-prompt or --append-system-prompt-file. Operates at system-prompt level. Must be passed every invocation, so better for scripts/automation than interactive use.
  • Need enforcement, not advice → CLAUDE.md is advisory. For actual blocking, use managed settings: permissions.deny, sandbox.enabled. CLAUDE.md = "do this", settings = "you cannot do that".
  • Policy goes in CLAUDE.md, enforcement goes in settings — official split:
Concern Where to configure
Block specific tools / paths Managed settings permissions.deny
Enforce sandbox isolation Managed settings sandbox.enabled
Auth method / org lock forceLoginMethod, forceLoginOrgUUID
Code style / quality guidelines Managed CLAUDE.md
Data handling / compliance reminders Managed CLAUDE.md

13. What's next

With CLAUDE.md solid:

  1. Claude Code install complete guide — if you haven't installed yet.
  2. Slash commands complete reference — /init, /memory, /compact, and everything else referenced above.
  3. Claude Code hooks complete guide — 4 lifecycle events × real scripts — CLAUDE.md is advisory; hooks are enforced (coming soon).

References

This is post 3/15 in the "AI Coding CLI Entry Guide" series. last verified: 2026-04-25 (per official Claude Code Memory + Best Practices docs).

댓글

댓글 쓰기

이 블로그의 인기 게시물

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

이미지
← 8/9 Deep Learning Architec… 📚 Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… 📚 Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere 핵심 요약 This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

이미지
← 7/9 Deep Learning Training 📚 Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. 핵심 요약 CNN : weight sharin...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

이미지
Series overview: Series index 📚 Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. 핵심 요약 An evaluation set for RAG is ...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks 📚 Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. 핵심 요약 SGD : \(w \leftarro...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent 📚 Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »

"ML Foundations (6/9) — Neural Networks: From Perceptron to MLP"

이미지
← 5/9 Regularization & Model… 📚 Series Index 7/9 Deep Learning Training → A neural network is logistic regression from Part 4 stacked into layers. Each layer passes through a new nonlinear activation, expanding the family of functions the network can represent. In this part we put how a neural network learns onto one page. Perceptrons, XOR, MLPs, backpropagation, activation functions, and the universal approximation theorem all meet here. 0. Learning Objectives Write the perceptron's learning rule and explain why it cannot solve XOR. Express an MLP's forward pass as one line of matrix products. Derive backpropagation by hand on a two-layer MLP. Compare the definitions, derivatives, and trade-offs of sigmoid, tanh, ReLU, and GELU. State the universal approximation theorem (Cybenko 1989 / Hornik 1991) precisely. Train an MLP end-to-end in PyTorch. 1. 핵심 요약 Perceptron (1958): \(\hat{y} = \mathrm{step}(w^\top x + b)\). Converges only on linearly separabl...
Read more »