CLAUDE.md Design Principles — Encoding an Agent's Identity in a Single File

스마트폰 속으로 들어온 24시간 AI 비서 — OpenClaw 텔레그램 연동 가이드

Global vs. Project Separation, Rule Conflict Prevention, and the "Prohibitions Over Instructions" Principle

Summary

  • CLAUDE.md is the agent's identity and behavioral contract. It loads automatically at every session start — what is written here defines everything the agent is.
  • The global file (~/.claude/CLAUDE.md) holds cross-project principles. The project-root CLAUDE.md holds project-specific rules.
  • Beyond 15 rules, contradictions become inevitable. Five clear prohibitions outperform ten ambiguous instructions.

What Is CLAUDE.md

When configuring an agent with Claude Code, CLAUDE.md is the first file you create. It is the agent's operational specification and the only "constitution" that is automatically read at every session start.

Claude Code loads CLAUDE.md in the following order at session initialization:

  1. ~/.claude/CLAUDE.md (global)
  2. CLAUDE.md at the project root

Both files are loaded into the context window. The agent has already processed this content before it receives its first message. The longer the files, the less effective context remains — and overlapping content wastes tokens.

Body

1. Global CLAUDE.md vs. Project CLAUDE.md — What Goes Where

The separation criterion is clear: "Does this apply identically across all projects?" → global. "Is this only meaningful in this project?" → project.

What belongs in global:

  • Project ontology — the full agent roster, roles, statuses, and relationships. An agent knowing "where I sit in the overall system" anchors its judgment.
  • User behavioral rules — conversation principles such as "present counterarguments first," "data over intuition," "respect autonomy." These must apply uniformly across all projects.
  • Inter-agent communication protocol — hub URLs, endpoints, payload formats. The interface for agent-to-agent communication must be consistent regardless of project.

Sharing the global file via symlink propagates any change to all projects immediately:

ln -s ~/.claude/CLAUDE.md /path/to/project/.claude/CLAUDE.md

What belongs in the project CLAUDE.md:

  • Project purpose — a single-sentence definition of what this agent does.
  • Tech stack and environment — OS, language, framework, build commands.
  • Workflow and pipeline — task flow, gate conditions, branching criteria.
  • File conventions — naming rules, directory structure, output format.
  • Project-specific prohibitions — constraints meaningful only in this project, such as "never generate figures without a source."
포럼 그룹을 활용한 토픽 분리

2. The 15-Rule Threshold — Why More Rules Cause Conflicts

Rules accumulate naturally. "This matters, and so does that..." The problem: beyond 15 rules, contradictory pairs become unavoidable.

Typical conflict patterns: - "Process user requests quickly" + "Run an interview for complex requests" → Should it be fast, or should it run an interview? - "Operate with minimal context" + "Read all relevant files thoroughly" → How much reading counts as thorough?

Humans resolve these through contextual inference; an LLM weights both rules equally. The result is inconsistent behavior — the agent follows a different rule each session.

Resolution: declare precedence explicitly, or constrain each rule's scope.

## Branching Criteria
- Simple (1–2 steps): orchestrator handles directly
- Complex (3+ steps): interview → research → compile pipeline
- When ambiguous, classify as complex (over- rather than under-classify)

Written this way, "fast" and "interview" no longer conflict. Simple → fast. Complex → interview. Ambiguous → complex. The decision tree is unambiguous.

The fastest way to find rule conflicts is to ask the agent directly: "Find any contradictory rule pairs in this CLAUDE.md." It surfaces conflicts that human reviewers miss.

중앙 라우팅 설정

3. The "Prohibitions Over Instructions" Principle

This is the most critical principle in CLAUDE.md authoring.

Ambiguous instruction:

Write at an appropriate length.

Clear prohibition:

Do not exceed 500 tokens.

Instructions leave room for interpretation — "appropriate" varies by context. Prohibitions are binary: violated or not.

Prohibition patterns that proved effective in practice:

Ambiguous Instruction Clear Prohibition
Cite sources properly Never generate a figure without a source
Respect the user's perspective Never generate AI viewpoints in the "My Take" section
Write concisely Never inline tool output exceeding 50 lines — save to file
Verify thoroughly Never use phrases like "probably fine"

Prohibitions outperform instructions because LLMs comply with "do not" more consistently than "do." Positive instructions introduce a degree problem; prohibitions only require a binary violation check.

4. Structural Design — Required vs. Omittable Sections

Sections required in a project CLAUDE.md:

1. Project purpose (2–3 lines)
2. Development environment (table)
3. Agent role separation (orchestrator / executor / quality)
4. Workflow pipeline (request flowchart)
5. Gate conditions (entry and exit criteria per phase)
6. File and naming conventions
7. Prohibitions (absolute rules)
8. Memory and context management strategy
9. Project structure (directory tree)

What each section does:

  • Agent role separation: Assigning orchestrator to dialogue and judgment, executor to implementation, and quality to verification draws clear role boundaries. Agents can self-determine "who should handle this."
  • Gate conditions: Specify Entry and Exit conditions for each phase. Be concrete: "The brainstorming phase exits only when the user approves the design." Without conditions, agents skip phases.
  • Memory and context management strategy: Defines how knowledge persists across sessions. Declaring a three-layer memory structure (project knowledge / behavioral rules / session restore) reduces unnecessary file traversal.

Sections that can be omitted: - General coding principles (agents already know these) - Detailed tool usage instructions (extract into skill files) - Frequently changing information (move to the memory layer)

5. Separating the Skill System from CLAUDE.md

Embedding procedural steps (step 1, step 2, step 3...) in CLAUDE.md loads them unnecessarily every session. Procedural task instructions belong in skill files.

.claude/
└── skills/
    ├── research/SKILL.md      # research procedure
    ├── compile/SKILL.md       # source file authoring procedure
    └── verify/SKILL.md        # fact-check procedure

CLAUDE.md retains only trigger keywords and skill paths:

## Skill Trigger System
| Keyword | Skill | Path |
|---|---|---|
| research, investigate | research | .claude/skills/research/SKILL.md |
| verify, fact-check | verify | .claude/skills/verify/SKILL.md |

This separation keeps CLAUDE.md to "what" and "why," while skills carry the "how." Context efficiency improves, and skills can be modified or reused independently.

6. Anti-Patterns — What Not to Do

Anti-pattern 1: The monolithic CLAUDE.md

Cramming all knowledge, all rules, and all examples into one file. Beyond 500 lines, agents begin discounting the latter sections. CLAUDE.md loads at the top of context, but excessive length pushes it down as the conversation grows.

Anti-pattern 2: Emotional instructions

This is really important, please make sure to follow it!!!

Exclamation marks and emphasis have no effect. Structural language such as "strictly prohibited" is more reliable.

Anti-pattern 3: Leaving conflicting rules unresolved

Writing "minimize token usage" alongside "read all relevant files" without providing a tiebreaker. The agent makes a different choice each session.

Anti-pattern 4: Conflating skills with rules

Inlining procedural task instructions into CLAUDE.md. Extracting them into skill files keeps CLAUDE.md lean and ensures skills are only loaded when needed.

Design Considerations

  • Starting without a global CLAUDE.md accumulates duplication. User preferences and communication protocols get rewritten per project, and any update requires manual propagation. Sharing the global file via symlink eliminates this.

  • Removing rules is harder than adding them. To determine whether a rule is necessary, observe the agent's behavior without it. Some rules have no effect when removed — the agent already follows them by default. Delete those.

  • Ask the agent to find rule conflicts. "Find any contradictory rule pairs in this CLAUDE.md" surfaces conflicts that human review misses.

  • Do not put volatile information in CLAUDE.md. API keys, version numbers, current status — these belong in the memory layer (tasks/sessions/, docs/). CLAUDE.md must contain only stable information.

Conclusion

CLAUDE.md is the agent's identity. Because it loads automatically every session, what is written here becomes the agent's decision framework.

A well-designed CLAUDE.md has few rules. Separate global from project, replace ambiguous instructions with clear prohibitions, extract procedures into skills, and move volatile information to memory. What remains is the core — nothing more.

"If an agent reads this file, does it fully understand its role?" — When the answer is yes, CLAUDE.md is complete.

댓글

댓글 쓰기

이 블로그의 인기 게시물

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 (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 »

"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 (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 »

"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 »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… 📚 Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
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 »