Coding Agents in Practice (1/5) — Claude Code Workflow: CLAUDE.md, Skills, Memory, Session Recovery

AI ์ฝ”๋”ฉ ์—์ด์ „ํŠธ ์‹ค์ „ (1/5) — Claude Code ์›Œํฌํ”Œ๋กœ์šฐ: CLAUDE.md·์Šคํ‚ฌ·๋ฉ”๋ชจ๋ฆฌ·์„ธ์…˜ ๋ณต์›

Same model, different workflow — different results. Four core building blocks, organized as a real-world pattern.

ํ•ต์‹ฌ ์š”์•ฝ

  • Four pieces: CLAUDE.md (identity) / Skills (task modules) / Memory (cross-session persistence) / Session Recovery
  • Primary source: Anthropic Claude Code documentation at code.claude.com
  • Core principle: "An agent is the model plus everything that sits on top of it" — and how that everything is structured is the workflow
  • Outcome: a well-shaped workflow reduces token cost, retry count, and debugging cost on the same task

1. Why workflow design now beats model choice

When coding agents were new, the answer was always "use a stronger model." By 2026, the same Sonnet 4.6 or Opus 4.7 produces meaningfully different results depending on workflow design. Stronger models demand more accurate context — and when context is wrong, a stronger model gives you wrong answers more confidently.

Claude Code provides four building blocks for shaping the workflow. This piece walks through each, with the underlying principle and the real-world pattern.

2. CLAUDE.md — Agent identity

Role. A markdown file at the project root. The agent reads it automatically at the start of every session. It carries "what is this project, and what are the rules?"

Authoring principles. - Read First section: list the additional core docs the agent should pull in. Keeps onboarding deterministic. - Hard Rules: spell out what must never happen — external publishes, credential edits, sweeping rewrites. - Workflow: a rule like "3+ steps → brainstorming → writing-plans → execution → verification" tells the agent how to escalate. - Scope / Surface: the directory structure and the responsibility of each directory.

Common mistakes. - Writing too long. CLAUDE.md is always in context; every line costs tokens. Keep it lean. - Forgetting "what does this project do?" The agent needs that on day one. - Embedding fast-changing info. Move volatile info to a separate docs/ file and link from CLAUDE.md.

Test. "If the next session starts fresh and reads only this CLAUDE.md, can the agent follow this project's rules?"

3. Skills — Task modularization

Role. Files at .claude/skills/<skill-name>/SKILL.md. Each skill is a reusable task package corresponding to one kind of task (e.g., publish, verify, git-commit, compile).

When to extract a skill. - The same kind of task happens repeatedly (blog publish, commits, meeting notes). - The task has multiple stages (validate → process → output → report). - The procedure is something the team should share.

Authoring principles. - Triggers: list the keywords that should activate the skill. - Stages: input → validation → processing → output → report. - Hard rules: things this skill must never do (e.g., "do not publish without user confirmation"). - Output format: how to present results to the user (markdown table, JSON, Discord message, etc.).

Avoid duplication. When two skills overlap, merge them. Too many skills confuses skill selection. 5–15 is a healthy operating range.

4. Memory — Cross-session persistence

Role. Information that should outlive any single session. Claude Code memory typically operates in three tiers:

  • User profile (memory/user_profile.md, feedback_*.md): the user's role, preferences, past corrections.
  • Project memory (memory/project_*.md): in-progress decisions, milestones, external system IDs.
  • References (memory/reference_*.md): pointers to where information lives outside the project (Linear project keys, Grafana dashboards).

Save vs. don't save. - Save: explicit user rules, corrections of bad approaches, non-obvious reasons behind decisions, external system coordinates. - Don't save: facts derivable from code, change history (git log is authoritative), one-shot task details, anything already in CLAUDE.md.

The MEMORY.md index. All memory enters through a MEMORY.md index file. The index is always loaded into context — keep entries under 150 characters per line and the file under 200 lines.

Validity check. Memory ages. Before recommending from memory, verify against the current state of the code or external system.

5. Session Recovery — Resuming safely

Problem. Claude Code sessions end — context limit, user closure, system restart. The agent needs a way to not lose where it was and resume from that point in the next session.

Three durable stores. - tasks/plan.md: current milestones, completed items, next action. The single read-first entry point at session start. - tasks/lessons.md: a what-worked / what-didn't / why table. Prevents repeating the same mistake. - tasks/sessions/YYYYMMDD-HHMMSS.md: one snapshot per session. At session end, log changed files, decisions, and next steps.

Session start hook. Claude Code's SessionStart hook can auto-inject these three. The result: a new session starts as if the prior context were already recovered.

Recovery cautions. - Verify that files or features the previous session claimed to create actually exist. Memory is a point-in-time observation, not live state. - Long-paused work may have external state drift (branches, dependencies, data). Run git status and dependency checks before resuming.

6. The integrated workflow — One real cycle

How the four pieces combine in a single cycle:

  1. User request → the agent recognizes a trigger keyword and calls a Skill
  2. Skill entry → check CLAUDE.md rules + memory + current state
  3. 3+ step task → pass through brainstorming → writing-plans → execution → verification gates
  4. Session end → log changed files, decisions, and next step into tasks/sessions/
  5. Next session start → plan.md / lessons.md / latest session.md auto-injected → resume immediately

Signs the cycle is working. - The user does not have to repeat the same answer twice (memory is working). - The user does not have to ask "where did we leave off?" (session recovery is working). - No external publish or destructive change happens without user confirmation (Hard Rules are working).

7. At a glance

Component Location Core line Common mistake
CLAUDE.md Project root Identity + Hard Rules + Workflow Too long / volatile info inside
Skills .claude/skills/<name>/SKILL.md Task module: triggers + stages + output Too many overlapping skills
Memory memory/MEMORY.md index + per-file Non-obvious cross-session info only Saving facts derivable from code
Session Recovery tasks/plan.md + lessons.md + sessions/ Where the next session resumes Trusting memory without verification

Next up

Part 2/5 lines up Cursor 3 vs Claude Code vs GitHub Copilot — May 2026 comparison. Three coding agents with different design philosophies, sorted by price, capability, and real workflow fit.

References

  • Anthropic, Claude Code Documentation — code.claude.com (verified 2026-05-05).
  • Anthropic, Hooks and Skills Reference — code.claude.com/docs/hooks (verified 2026-05-05).
  • The "What Is Harness Engineering?" series provides the theoretical background for workflow design.

This is part 1/5 of the Coding Agents in Practice series.

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

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

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ 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"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š 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"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š 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. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

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

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š 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. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

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

← 6/9 Neural Networks ๐Ÿ“š 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. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

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

← 1/13 Current Structure Inve… ๐Ÿ“š 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 ๐Ÿ“š 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 »