AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people who "build an AI team" build one big chat with a long prompt and call it a team. It feels organised because the prompt has seven job titles in it. In practice every job shares one context window, one set of instructions and one pile of half-remembered corrections, so the research bleeds into the drafting and nothing runs unless you turn up.
The rule: an agent is only real when it has four things written down — a single job, its own folder, its own memory file and its own trigger. Anything missing is a role-play, not a team member.
A note on the source idea. This topic comes from a creator describing seven named Claude agents, each with its own folder, memory and schedule. I could not verify that setup, the names, or what it produces, so nothing below depends on it. What is verifiable is the plumbing Claude actually offers for each of the four things, and that is what this guide uses. Names are your choice; they are labels, not features.
You need Claude Code (terminal, desktop app or web) and a folder for the business. Claude Code is included in paid Claude plans. At time of writing, Pro is listed at $20 a month (about £15 at time of writing) and Max starts at $100 a month (about £75 at time of writing). Check the £ price at checkout, because it moves with the exchange rate and plan changes. Scheduled cloud work draws on the same usage allowance as normal sessions, so more agents means faster use of it.
One card per agent, kept in a single ROSTER.md at the top of your business folder. Each card has six lines.
| Line | What you write | Where it lives in Claude Code |
|---|---|---|
| Name and job | One verb phrase, e.g. "Triage the support inbox" | name and description in the agent file |
| Folder | What it may read and write | A subfolder, plus the tools list in the agent file |
| Memory | What it must remember between runs | memory: field, or a notes file it maintains |
| Trigger | Manual, scheduled or event-driven | A routine, desktop task or /loop |
| Output | Where the result lands and what a good result looks like | A named file or a draft for you |
| Stop rule | What it must never do without you | Permission mode and prompt constraints |
If the sentence needs "and", split it into two cards. "Research competitors and draft the newsletter" is two agents with a handover file between them. The description field in a subagent file is what Claude reads to decide when to delegate, so a vague job also means vague delegation.
Create one subfolder per agent, for example agents-work/inbox/, and a shared handover/ folder. The agent reads what is in its folder and writes its result there. Handovers between agents go through files you can inspect, not through hidden chat context. Limit the agent's reach with the tools field in its definition; leaving tools out means it inherits everything, which is the wrong default for a helper.
Claude Code subagents are Markdown files with YAML frontmatter. Project agents live in .claude/agents/; personal ones in ~/.claude/agents/. Only name and description are required. Adding memory: project gives that agent its own persistent directory at .claude/agent-memory/<agent-name>/, and the first 200 lines or 25KB of its MEMORY.md is loaded into its prompt. The other scopes are user (~/.claude/agent-memory/<name>/, across all projects) and local (.claude/agent-memory-local/<name>/, not for version control).
A minimal agent file:
---
name: inbox-triage
description: Sorts new support emails into urgent, routine and ignore, and drafts replies for routine ones. Use when asked to triage the inbox.
tools: Read, Write, Grep
model: sonnet
memory: project
---
You triage support email. Follow the rules in your memory file.
Never send anything; write drafts to handover/inbox-drafts.md.
When I correct you, record the correction in memory.
Keep the shared CLAUDE.md for rules everyone follows. Anthropic's guidance is to aim for under 200 lines per file, because longer files reduce how reliably instructions are followed. Put agent-specific lessons in that agent's memory, not in the shared file.
Pick the trigger that matches where the work needs to run. These are the three Claude Code options as documented:
| Option | Runs on | Needs your machine on | Minimum interval |
|---|---|---|---|
| Routines (cloud) | Anthropic cloud, fresh clone of your repo | No | 1 hour |
| Desktop scheduled task | Your machine, local files | Yes | 1 minute |
/loop |
Your open session | Yes, and session open | 1 minute |
Catches worth knowing. /loop recurring tasks expire after seven days. Routines are in research preview, so limits may change. They run autonomously without permission prompts, so scope their repositories and connectors tightly, and they belong to your own account rather than a team. A cloud routine works from a fresh clone, so local-only files are not there. Times set exactly on the hour can start late; choose a minute past, such as 9:07. Create a routine with /schedule in the CLI or at claude.ai/code/routines. Routines are available on Pro, Max, Team and Enterprise.
Write what a good result looks like in the agent's prompt: format, length, where it saves. Then write the stop rule: drafts only, no sending, no deleting, no spending. For anything that touches money, customers or public posts, the output is a draft in a file and you press the button.
Imagine a small UK candle shop. This is invented for illustration. Three cards, not seven.
| Agent | Job | Folder | Trigger | Output |
|---|---|---|---|---|
| Orders | Summarise yesterday's order export and flag anomalies | orders/ |
Desktop task, weekdays 07:07 | handover/orders-summary.md |
| Inbox | Triage support email, draft routine replies | inbox/ |
Manual, when you sit down | handover/inbox-drafts.md |
| Writer | Draft one newsletter from the other two files | writer/ |
Manual, Fridays | writer/draft.md |
The Writer reads only the handover files, so you can see exactly what shaped each draft. Each agent's MEMORY.md collects its own corrections: the Inbox agent learns which suppliers to ignore, the Writer learns the shop's tone. After a month, review each memory file by hand and delete what is wrong.