Chapter 4

Create an environment where AI can work

Let AI retrieve relevant information and continue the work

The salesperson spends the evening collecting files for AI. At 9 p.m., he is exhausted and still trying to remember where a document is stored.

Once you can delegate work, another problem appears: collecting the Goal and documents for each request takes 30 minutes. This chapter removes that repeated preparation.

The aim is to let AI retrieve the information it needs independently. That is what an AI workspace provides. AI can also find relevant sources you forgot to include, improving the quality of its work.

In the Goal Loop, the effort required to delegate Step 3 changes. Instead of collecting information for every ToDo, you keep it accessible. Step 4, updating the current state, also happens as AI works. Steps 1 and 2, choosing the main constraint and the ToDo, remain your responsibility.

Capability alone is insufficient without knowledge of the workplace. AI needs the locations of shared folders, the history of decisions and an understanding of who decides what. Providing that orientation once can make AI useful without weeks of preparation.

Keep environment setup simple

Discussions of AI environments often introduce RAG, vector databases, knowledge graphs, agent frameworks, memory and evaluation systems. Studying the terms can consume an entire day.

They are names for methods. Learning the names does not create a working environment. The essential requirement is one sentence:

Every company can use the same AI models. You and your competitors can access the same underlying capabilities. The difference is what you make available to those models.

AI capability is available to everyone. Your working environment depends on you.

Your sales records, decision history and current work state exist in your company. AI becomes useful when it can read that information.

Four steps to prepare the workspace

There are four steps. A one-time, 30-minute setup can replace 30 minutes of preparation for every request.

1. Identify where the information lives

List the locations of information used in your work. Most fit into six categories.

Do not try to inventory all company information immediately. That quickly becomes difficult to sustain. Start with the information needed for one action you want to delegate. Preparing course materials may require only your computer and a shared drive. Planning next year's business may require financial spreadsheets and customer databases too. Add sources as the scope of delegated work grows.

The last category, knowledge that only you hold, is easy to overlook.

The salesperson realizes that only he knows the customer's real concern. He writes a note in the project folder, making that context available for AI to read.

“The customer cares more about the implementation date than the price” may not be written anywhere. AI cannot know that fact until you record it.

2. Put the sources where AI can read them

Create a folder for each project and place the relevant sources in it. Keep the structure simple.

projects/
  acme-training/
    01-meeting-notes.md        # Original meeting notes
    02-proposal-v3.pptx        # The version actually sent
    03-quotation.xlsx
    customer-context.md       # Knowledge previously held only by you
  sales-hiring/
    job-description.docx
    interview-notes-2026-08.md

Now AI has a clear place to open when you mention Acme. Follow three rules.

Use one question: “Would I open this source to do the job?” If the answer is no, leave it out.

Organization becomes useful when AI can access the sources. That is the next step.

3. Connect the sources

Once the files are organized, make them readable from the AI workspace.

An AI working with local files can read, write and create them. It does not automatically have access to send email or add calendar events. Connections to external tools provide those capabilities.

MCP, connectors, CLI and API are different ways to provide access to tools. You do not need to memorize every service that can be connected. Learn how to connect and disconnect a tool, then choose the connections needed for the work.

There are three practical categories:

1. Project folders: tell AI where the project's information is stored.
2. Email and calendars: provide access to correspondence and schedules.
3. Business tools: connect the systems containing customer, accounting or other operational data.

If you are unsure what to connect, ask AI:

Tell me which missing data access has prevented you from completing the work.

AI can identify a useful connection. Add it when the need appears. Ask both what should be connected and how to connect it. The video in Supplementary lesson 1 demonstrates the process.

The salesperson follows those steps.

The salesperson spends 30 minutes organizing existing documents by project and connecting the folder to AI. He does not need to create replacement documents.

4. Decide access and review boundaries

Decide what AI should not see. You make that decision yourself.

Exclude confidential information entrusted to you by other people when you are not authorized to share it: personnel evaluations, health information and confidential customer documents. Do not place material in the workspace if you cannot independently authorize its disclosure.

Also define review requirements for information leaving the organization. Customer emails, external sharing and public posts must be reviewed by a person before sending. AI can read and prepare the work, while you retain the final send action.

This step concerns two practical decisions: which folders AI can access, and which actions require human review.

Add one instruction file for AI

After organizing and connecting the sources, prepare one short instruction file. List where information is stored, one location per line. Do not copy the contents into the file. Explicit locations tell AI where to retrieve the sources.

The example uses Addness, a tool where people and AI can read the same Goals and current work state. Documents and chats can remain in their existing tools. Addness provides a shared place for what the team is trying to achieve and how far the work has progressed. Chapter 6 explains why that matters.

The instruction file has a tool-specific name: CLAUDE.md for Claude Code and AGENTS.md for Codex. AI reads it when starting work in the folder. A short file is enough.

# About the workspace
Owner: Sato, Sales (corporate training proposals)

## Information locations
- Project documents: projects/<project-name>/
- Daily notes: days/
- Decisions and current work state: Addness, the authoritative record

## Working rules
- Record progress and decisions in the relevant Addness Goal.
- Show customer emails to the owner before sending.

Keep detailed work procedures out of this file. A long file has to be read on every run and can slow work down. Limit it to source locations and working rules.

You do not need to write the file before trying AI. Start in an empty folder and ask AI to set up a workspace for the job. Recent models can prepare the instruction file as part of that setup. Pay closer attention to the structure when you begin repeating the work or automating it, because those uses require more consistency.

Turn repeated explanations into a procedure

Repeated work reveals repeated instructions: “Write meeting notes in the order decisions, action items, next meeting,” or “Use the quotation template and show tax separately.” Save recurring procedures in a file instead of explaining them every time.

Claude Code supports Skills. A procedure stored in the designated location can be read when it is relevant. Ask AI to create the Skill, including placing it correctly.

Use the same process for future meeting notes.
Save the procedure from the completed example as a Skill.

Ask AI to create the procedure after a successful attempt. The steps that actually worked become the reusable method. Corrections are preserved in a file so later work can benefit from them.

Specific connection procedures change, while the structure stays useful. Tools change; the underlying approach remains. For instructions on installing an AI app and connecting it to Addness in your language, use the setup guide and A complete guide to Claude Code.

A prepared environment helps every later request

The salesperson's requests change after the setup.

The collection and transfer are no longer needed. You still decide what should be done, as described in Chapter 3.

After organizing the project folder and connecting AI, the salesperson can start work with one short request. AI retrieves the relevant material independently.

The activity log shows AI reading email, the project folder, instructor availability, the calendar and the customer-context note. It returns an instructor schedule proposal, a warning that preparation time is shorter than usual, and a draft email to the instructor. A person decides whether to send it.

The Goal and sources were already available. Repeated explanations disappear, and the effort required to delegate approaches zero.

The information also stays current as work proceeds. The instruction file says to return results to the Goal. Each job adds findings, meeting notes and decisions to the appropriate record. The next request begins from that updated state. You do not have to manually rebuild the information set each time.

A refined prompt helps one request. A working environment helps every request that follows.

At Addness, customer data integration work was passed from one AI to another without a separate handoff explanation. A handoff that had taken a full day took 30 seconds, because the next AI could begin from context stored in the Goal. The example on note describes the work.

Review

• The essential requirement is access to your work information.
• Everyone can use the same AI capability. Your environment depends on you.
• Follow four steps: identify sources, organize readable originals, connect them, and set access and review boundaries.
• Reuse existing documents. Write down context that only you know.
• Organization produces value when AI can use the sources.
• Decide what AI must not see and what a person must review.
• Add one short instruction file containing source locations and working rules.
• AI updates the work record as it acts, making the latest state available to the next request.
• In the Goal Loop, the changes are less preparation for Step 3 and automatic recording for Step 4. You still choose the constraint and ToDo.

Continue to the next chapter

You now have an environment where AI can work and participate in the Goal Loop.

The next chapter covers automating repeated procedures and letting AI run the Goal Loop autonomously.

• Chapter 5: Automate work and let AI manage decisions