Your First Autonomous AI Agent Session: The 15-Minute Setup
Everything on this site — the memory system, the contract, the guardrails — converges on one moment: the first session where your agent works alone and hands back finished work. This is the walkthrough of that moment: fifteen minutes of setup, in five steps, with the exact prompts. No programming anywhere.
Prerequisites: a file-based agent installed (Claude Code or equivalent), git installed (git-scm.com, default options — 5 minutes, once), and the free core templates.
Step 1: Build the office (3 minutes)
Create a folder — this is your employee's entire world. Name it after the venture, not the tool: newsletter-business/, not ai-test/. (You'll take it more seriously, and oddly, so will the agent — the folder name is context too.) Copy the template pack's core-system/ contents into it: eight files and a log/ folder. Done.
Step 2: Write the mission (5 minutes — the step that matters most)
Open GOALS.md and replace the placeholders. Two rules:
- Be honest, not impressive. Nobody's watching. "Figure out whether my cake-decorating knowledge can become a $500/month digital product" is a great mission. "Build a revenue-generating content empire" is a costume.
- Make success measurable. The 30/90/365-day slots force this. If you can't measure it, the agent can't know if it's working — and honestly, neither can you.
Fill the constraints while you're there: budget (exact — "zero, free tools only" is a fine answer that will be respected), your availability ("~10 min/day"), anything off-limits. Then open BACKLOG.md and write task T-1. If the mission is clear but the path isn't, use the universal opener:
T-1: Research how to best accomplish the mission in GOALS.md. Propose a concrete plan as new backlog tasks, with reasoning in DECISIONS.md.
This task is special: its output is more tasks. You're not planning the work — you're delegating the planning.
Step 3: Sign the contract (2 minutes)
Read CLAUDE.md top to bottom, as a manager reads a contract before signing — you're checking for anything you don't actually want. Adjust the guardrail list to what genuinely scares you (guidance here). Two pages. When it reads true, it's signed.
Step 4: The first session (3 minutes of yours; hours of theirs)
Open your agent in the folder and say:
"Read CLAUDE.md and begin. First: initialize a git repository and make the first commit. Then take T-1."
Now watch — first days deserve supervision. Within minutes you should see the pattern this whole system promises: files read, mission confirmed briefly (a paragraph, not an interrogation), git initialized, work starting on T-1. Within the hour: research happening, notes accumulating in a project folder — and if anything touches the guardrails, a tidy proposal in FOR_HUMAN.md instead of a unilateral action. Let it run.
Step 5: The inspection (5 minutes — and only this once)
After the session, one full walkthrough. Not because you'll do this daily (you won't) — because seeing the machinery work once is what lets you trust it after:
STATE.md— does it say precisely where work stopped and what's next? Could a stranger continue from it?log/— is there today's entry, and does "what failed" contain anything? (An honest log has failures. A spotless one is a red flag, not a gold star.)FOR_HUMAN.md— anything waiting? Answer it now, in writing, under the item.git log --oneline(ask the agent to show you) — a timeline with meaningful notes?- Anything off — state too vague, log missing? Say so, then have the agent encode the correction into CLAUDE.md. Day one is when contract-tuning habits form.
Tomorrow: the payoff
Open a fresh session — the amnesiac, remember — and type the entire briefing:
"continue"
Watch it reconstruct: files read, context rebuilt, work resumed exactly where it stopped, no questions the files already answer. The first time a stranger picks up yesterday's work in two minutes flat is the moment this stops being an article you read and becomes a thing you own. From there, your job is ten minutes a day — and the agent's job is everything else.