The AI Agent Playbook: Notma Field Notes

01

Chat vs. agents: cross the chasm

The playbook's core distinction is simple: chat is question to answer; agents are goal to result. A chat model gives you information, then you still do the work. An agent takes a defined goal, breaks it into steps, uses tools, and hands back an output.

For Notma, this is the practical line between AI literacy and AI integration. A business has not adopted agents because staff can ask better questions. Adoption starts when meaningful work actually leaves the team's plate.

• Chat input: a question. Agent input: a goal.
• Chat output: an answer. Agent output: a finished result.
• Chat is useful for lookups and simple decisions. Agents are useful for multi-step work.

02

The agent loop: observe, think, act

The playbook frames agents as a loop: observe the current state, think about the next best action, act through a tool or output, then observe again. The loop repeats until the agent decides the goal has been met.

That makes precision more important than magic. If the task says 'research podcasts and make a deck,' the agent needs enough detail to know what counts as done: how many podcasts, which format, what the deck should contain, and what quality bar it must hit.

• Observe: gather current state, files, tools, and data.
• Think: choose the next best step.
• Act: search, read, write, call a tool, launch a subtask, or produce output.
• Loop: continue until the task has a clear finish line.

03

Onboard the hire: context, tools, skills

A raw agent is a capable stranger. It may be smart, but without your business context, your preferred tools, and your operating procedures, it will improvise. The playbook reduces agent onboarding to three levers: context, tools, and skills.

This is the same adoption problem Notma sees in businesses. A team does not need a mystery box; it needs a system that knows the business, can reach the right tools, and follows explicit operating rules.

• Context: what the agent should already know before it receives a task.
• Tools: what the agent can use to take action in the real world.
• Skills: repeatable SOPs that define how output should be produced.

04

Own your context files

The playbook argues for business context in plain Markdown files rather than only inside platform memory. This makes context portable, inspectable, and editable. An agent can read the files at the start of a session and work from the same source of truth every time.

The recommended stack includes a master file, personal or founder background, business information, brand voice, ideal customer profile, offer catalog, and memory. The important point is not the exact file names; it is ownership and clarity.

• claude.md or agents.md: the north-star file that points to the rest of the system.
• about_me.md: personal background and founder story.
• business_info.md: what the business does, who it serves, and how it makes money.
• brand_voice.md: tone, language rules, and examples.
• ideal_customer_profile.md: who the business targets and who it does not.
• offer_catalog.md: services, pricing, upsells, downsells, and constraints.
• memory.md: lessons, preferences, and recurring patterns from past work.

05

Use a master file as the agent's north star

The master file is the agent's briefing document. It loads at the start of a new session and tells the agent where the context folder, skills folder, and available tools live. Done well, it reduces context rot because a fresh session can start informed instead of inheriting a long messy chat.

The playbook suggests three levels: one master file only; a master file plus context folder; and a full operating system with context, skills, and tools. Notma's recommendation is to start with the smallest version that survives real work, then expand when the team keeps repeating the same instructions.

• V1: one master file with the minimum viable business briefing.
• V2: master file plus a context folder for brand, ICP, offers, and memory.
• V3: master file plus context, skills, tools, and department folders.

06

Turn repeated instructions into skills

A skill is an SOP for an agent. If you repeatedly explain the same output style, report format, hook structure, email pattern, or review step, that repeated instruction belongs in a skill file.

The playbook's useful test is this: every time you re-explain something to an agent, you have discovered a skill that should be written down.

• Identify a repeatable output or process.
• Document the exact process in Markdown with examples.
• Store it in a skills folder with a clear name.
• Point the master file to the skills folder.
• Update the skill when output drifts.

07

Tools and MCPs expand the act step

Tools are what let an agent take action: search the web, read and write files, send email, post to Slack, pull product information, schedule campaigns, or upload creatives. Without tools, the agent can advise. With tools, it can run a workflow.

The playbook highlights MCPs as one way to connect agents to external services. The deeper lesson for a business is that each integration should be tied to a workflow and a review point, not added because it sounds powerful.

• Connect tools only where the agent needs to act.
• Give the agent business context before granting action tools.
• Use human approval gates for consequential work.
• Treat each new integration as part of an operating system, not a toy.

08

The Agent OS folder structure

The playbook's operating-system idea is a structured project folder that an agent can live inside. It is modular: each department or project can have its own context and skills, while the master file tells the agent how to navigate the system.

For business owners, this turns AI setup into a company asset. You can move the folder between agent tools, inspect the logic, and improve the system over time instead of starting over on every platform.

• /project-root: the top-level folder opened in the agent harness.
• claude.md: the orchestrator that auto-loads on every session.
• /context: background files such as about, business info, brand voice, ICP, offer catalog, and memory.
• /skills: process SOPs such as title hooks, email format, ad copy, negotiation, report layouts, or research workflows.
• /department: optional folders for marketing, sales, finance, operations, or brand work.

Notma note

This is a Notma field-note summary of a third-party playbook, not a verbatim republication. The goal is to make the operating ideas useful for founders and teams building practical agent workflows.

Open the source document on Google Drive →