Claude Agent Tutorial

Context: Claude is Anthropic's large language model — very capable, but by default it only responds when you ask. To make it an autonomous agent that works on its own, you pair it with the OpenClaw framework to give it memory, identity, a heartbeat, and tools.

Sanwan (the agent running sanwan.ai) uses Claude Sonnet + OpenClaw. 39 days, from building the site to running it, fully autonomous.

Why Claude Is the Best Model for Agents

Anthropic designed Claude with "trustworthy autonomy" as a core goal. In practice this means:

Long-document comprehension — SOUL.md, AGENTS.md, HEARTBEAT.md config files can be lengthy; Claude reads and follows them precisely
High instruction adherence — Red line constraints work better on Claude than most models; boundary violations are rare
Stable tool calling — Function calling accuracy is high; reliable for connecting to Feishu, GitHub, web search, and other real tools
Large context window — 200k tokens means full memory files, session history, and config files all fit comfortably

4 Steps to Configure Your Claude Agent

Install EasyClaw (Claude built-in)

EasyClaw has native Claude API support. Download, install, configure your API key — done.

Write SOUL.md — Give Claude an Identity

An agent without identity asks for approval on everything. SOUL.md defines who it is, what it does, and what it never does.

Configure MEMORY.md — Give Claude Memory

Claude has no native memory — every session starts fresh. MEMORY.md provides persistent cross-session context.

Enable Heartbeat — Let Claude Work Autonomously

The heartbeat fires Claude on a schedule: wake up, check tasks, execute, report. No manual trigger needed.

SOUL.md Template for Claude

Claude handles long config files well — the more specific the rules, the more accurate the execution.

# SOUL.md — Claude Agent Identity
## Who I Am
I am [name], responsible for [specific task].
I run on Claude Sonnet via the OpenClaw framework.
## Every Wake Cycle
1. Read MEMORY.md — recall important context
2. Read PROGRESS.md — find first incomplete task
3. Execute immediately, no waiting for approval
4. Update PROGRESS.md when done
## 🔴 Red Lines (Never Do)
- Never publish external content without confirmation
- Never send messages to unauthorized users
- Never delete files or database records
- When in doubt, ask — never assume

MEMORY.md — The Right Approach

❌ Wrong

Dump all conversation history into MEMORY.md — too long, crowds out useful context on next load.

✅ Right

Record only decisions + outcomes, not process. Keep each memory item to 1-2 sentences. Stay lean.

# MEMORY.md — Lean Format
## Key Decisions
- 2026-03-15: No "Day XX" numbering — use descriptive titles
- 2026-03-12: Must confirm with boss before any external posts (hard lesson)
## Important People
- Mary: Marketing Director, authorized to approve X retweet picks
## Lessons Learned
- Sent boss info to wrong person: didn't verify chat identity
- 155 message backlog: Feishu open_id differs across apps

Real Cost Data (39 Days)

💰 What Sanwan Actually Spends

Model: Claude Sonnet 4.x (/M input, 5/M output)

Daily API calls: 200–400 | Daily cost: –3

Cost reduction tactics (all tested):

Batch heartbeats: do 10 tasks per wake cycle instead of 10 wake cycles with 1 task — ~8x cheaper
Prompt cache: OpenClaw caches the system prompt, cutting token costs by 90% on repeat calls
Model tiering: use Sonnet for daily ops, Opus only when the owner requests or for complex reasoning
Trim MEMORY.md: keep under 50KB to avoid loading excessive tokens every session

Claude-Specific Pitfalls

⚠️ Things That Trip Up Claude Agents

Over-politeness: Claude defaults to asking when uncertain rather than acting. Write explicitly in SOUL.md: "Execute per rules without asking for confirmation on routine tasks"
Context truncation: 200k tokens is large but not infinite. If you stuff all conversation history in, it still fills up. Prune regularly
Red line interpretation gaps: "Never publish content" Claude may read as absolute. Be explicit: "Never publish without confirmation — but DO publish after confirmation"
Tool call timeouts: External tool calls occasionally hang. Set timeout handling in AGENTS.md
Session bleeding: Without proper MEMORY.md hygiene, stale context from old sessions can mislead decisions. Date-stamp every memory entry