AI Sprint: OpenClaw — Session 2 (Token Anatomy, Claude Code Inside the Server, Skills, and Multi-Agent Workspaces)

The drains, in order of surprise: SYSTEM PROMPT TAX — every model call re-ships the agent's whole self-description (all the documents: soul, agent, tools, heartbeat, memory — '23 sections injected every turn'); his analogy: answering one viva question by re-reading the entire project handbook first. TOOL SCHEMAS — every tool call carries descriptions, parameters, and usage rules (~8k as-heard). CONTEXT REPLAY — conversation N re-sends 1..N-1: 'first message small; fourth message carries the first, second, and third.' HEARTBEATS — 'silent killers': his weather heartbeat burned ~170k tokens delivering... a reminder; he hit his limit without sending a single message. HIDDEN BACKGROUND CALLS — the invisible multiplier: title generation, tagging, follow-up suggestions, autocomplete; one visible reply, ~5 API calls. SUB-AGENT SPAWNS — every spawned agent is a NEW context paying the full tax again ('you have 10, 20 agents — imagine'). LARGE TOOL OUTPUTS — web pages, JSON dumps, email threads, calendar exports.

The economics punchline reframes the whole hype cycle: 'I don't think OpenClaw is making money. The model providers are making money.' The runtime's intelligence — context-gathering, orchestration, proactivity — IS the cost; you can't remove it, only manage it.

The countermeasures, mapped one-to-one against the drains: (1) START FRESH OFTEN — prepare what you want, ask in 1-2 messages, type /new in Telegram; memory.md carries continuity so the context doesn't have to. (2) SOUL.MD DIET — every character is sent every call; target under ~5,000 characters, audit for fat ('obesity is also a problem'). (3) HEARTBEAT ISOLATION — isolated session, tiny output, cheap model: ~170k → ~2k as-heard. (4) HAIKU FOR SIMPLE TASKS — ~80% saved on questions and drafts. (5) SMART ROUTING WITH FALLBACKS — chains, not single choices: when 4.1 hits a limit, auto-fall to 4o-mini; configured by prompting Claude Code. (6) PRUNE MEMORY.MD WEEKLY — 'memory.md is injected every call; outdated facts cost real money' (his Nemory automates exactly this). (7) BACKGROUND TASKS TO CHEAP MODELS — the unavoidable orchestration calls don't need premium reasoning: ~30% instant. (8) PROMPT CACHING — identical system prompts shouldn't re-bill: 'Anthropic charges 90% less cached.'

The delivery mechanism matters as much as the list: these eight headings become PROMPTS — 'give these headings to your Claude Code and tell it: look out for this in my OpenClaw configuration.' The audit is delegated to the same agent class that caused the bill.

The de-scaring sequence for non-tech people: a TERMINAL is 'ordering pizza by phone' — typing commands is the phone call, the VPS chef does the cooking; it's 'typing like SMS, but for computers.' Two commands are the whole entry ritual: SSH root@<VPS-IP> — the building (VPS), its address (IP), the gate (SSH), and the special access (root), password demanded at the door — then cd <docker-openclaw-project> to walk to the agent's room. Both are COPIED, never memorized: Hostinger's h-panel shows the root-access string and the Docker manager shows the project name ('I copy and paste — I'm very bad at memory; that's why I built a memory tool'). Then the move that changes everything: INSTALL CLAUDE CODE INSIDE THE SERVER (one npm-style command from the guide, login via browser auth) — from then on, configuration is conversation: restart the gateway, add models, wire fallbacks, install skills, all by asking.

The beginner loop wrapping it: a STARTER PROMPT makes ChatGPT a 'patient technical partner' ('I'm a beginner and non-tech... one safe step at a time; before each command ask me what I'm trying to do and what screen I'm on'); every terminal output goes back as a screenshot for the next instruction. The live ROOKIE MISTAKE seals the lesson: he ran a server command in local PowerShell, screenshotted the error, and ChatGPT diagnosed it — 'you are not on VPS right now' — and even apologized for not saying which machine. 'This is how you use AI. You make AI dance; AI shouldn't make you dance.'

The anatomy: a skill is a FOLDER — SKILL.md (plain text: what it does, when to use it, steps to follow) plus optional scripts and config. His teaching method is dissection: download a ClawHub zip, open every file, 'dig into the dish' like his aunt deconstructing restaurant food. Sourcing: the awesome-openclaw-skills GITHUB LIST (thousands — coding agents, browser automation, nano-banana image skills, voice via ElevenLabs, calendar, sheets, scraping, security audits) and ClawHub with its VirusTotal parameters checked — against the standing security memory of the marketplace malware incident. Installing is conversational: give Claude Code the skill name or repo link — the live FIRECRAWL install produced a full proposal (skill.md, README, Python script reading the API key from .env), waited for approval, wrote everything, and specified exactly where the key goes; his standing guardrails ride along: 'do not change anything unrelated in the project,' keep it simple, ask before creating files.

The DIY doctrine stays primary: marketplace skills encode SOMEONE ELSE's way of working — 'a custom skill turns your repeated prompt into reusable capital,' so write your own for your niche (the Day 1 daily-planner proved it needs no code). The GOOGLE CALENDAR skill is the tier-2 reality check: added via marketplace + API keys, demoed with genuine sync flakiness (4 AM debugging in his own chat history), a hallucinated date he catches live, AND the correct commitment behavior — 'this changes your calendar and creates a commitment. Is this correct?'

The architecture: multi-agent in OpenClaw is NOT many installs — it's one runtime hosting multiple AGENT WORKSPACES, each a folder with its own document set (soul.md, agent/identity.md, user.md, memory.md) and its own skills. The org chart: a LEAD AGENT (his is 'Duma' — DC name) receives every request, 'understands and decides,' does what it can, and DELEGATES the rest — agent 1 plans and prioritizes, agent 2 handles reminders, follow-ups, and tracking. Folder structure mirrors the hierarchy: main workspace on top, agent-1/agent-2 folders beneath, each fully documented. Built the same way as everything today: ask ChatGPT/Claude for the prompt, hand it to Claude Code, workspaces appear; skills are ASSIGNED per agent and movable ('the Google Sheets skill landed in the wrong agent — tell Claude Code to move it').

The economics carry Day 2's theme to its conclusion: each agent is a fresh context paying the full seven-drain tax — 'if you're running 3-4 agents, imagine the consumption... people who say 50 agents: boss, how are you taking care of your token consumption?' The survival move: per-agent MODEL assignment (reasoning agent on a premium model, conversational agents on Haiku/local). And the scope note that keeps it honest: 1-click cannot do this — multi-agent is precisely why self-managed exists.