Key points
- The framework has six steps: installation and a sandbox, a knowledge hub, a guide file for the agent, a report-only warm-up task, a first gated workflow, and a weekly rhythm. A week to roll it out is a work plan, not a promise of business results.
- You set up safety before the first command: the agent works in a separate folder on copies of your data, originals are untouchable, and passwords and keys never end up in working files.
- The knowledge hub is one folder of plain text files: company, brand, products, processes, decisions. What lives in heads and email threads today goes into files. Customer personal data, passwords and confidential contracts never do.
- The system keeps getting better through the correction loop: every recurring review note gets written into the knowledge hub, so the agent never makes the same mistake twice.
The previous chapters answered one question: what AI can do in your store and how to supervise it. This bonus answers a different one: what to do on Monday morning. You get a ready framework in six steps. Each step fits into an afternoon, most into an hour, and you can realistically roll out the whole thing in a week.
One caveat up front: a week is a work plan, not a promise of business results. After seven days you have a working system and first outputs on the table, not a guaranteed sales lift. You do not need a developer or an implementation budget. You need a computer, one folder and a bit of discipline.
Step 1: Install Claude Code and give it a safe sandbox
Claude Code is an AI assistant that runs in the terminal, directly on your files. It reads them, transforms them and reports on them, and you give instructions in plain language. The terminal looks austere, but that is an illusion: you write to the agent the way you would write to an employee. Installation and first launch are covered step by step in our Claude Code guide, so we leave that part there instead of repeating it here.
Before you issue the first command, prepare the sandbox. Create a separate working folder, for example ai-work, away from the files your store actually runs on. Everything the agent is meant to read goes in there as a copy: the catalog export, the price list, the templates. Two rules apply from day one, with no exceptions:
- Originals are untouchable: the agent works on copies only.
- Secrets stay out of working files: passwords, API keys and login credentials never go into them.
These two rules cost five minutes of attention and remove most of the real risk from the entire rollout.
Step 2: Build the knowledge hub
The knowledge hub is one folder of plain text files. No new system, no licenses, no IT project. Plain markdown is enough, because the agent reads text files directly and needs nothing more. This is the most important step of the whole framework: the agent is exactly as good as the knowledge you give it to read. The starter structure looks like this:
- company/: what you sell, to whom and through which channels, plus a glossary of company terms (what "bestseller", "season" or "master carton" means in your business).
- brand/: tone of voice with examples of good and bad descriptions, language rules, claims that must never be used.
- products/: data rules, meaning units, dimension formats, the color dictionary and required fields per sales channel.
- processes/: how you handle an order, a return and a complaint, who owns what and when an issue escalates.
- decisions/: a decision log with dates and reasoning, so knowledge does not vanish when someone leaves or forgets.
The filling rule is simple: what lives in heads and email threads today goes into files. Every answer you give for the second time is a candidate for writing down.
Just as important is what does not go into the knowledge hub: customer personal data, passwords and confidential contracts. The agent does not need them, and you do not want them in a folder a machine reads. We say this with conviction because this is exactly the model we run our own company and our client rollouts on: our agents start every session by reading the company knowledge hub.
Step 3: Write a guide file for the agent
In the main folder, place a single instruction file, for example README-AGENT.md. It is the first thing the agent reads in every session, so treat it like a new-hire onboarding compressed to one page. Three sections are enough:
- Who we are: two or three sentences about the company and the assortment.
- Where things live: a map of the knowledge hub folders with one line of description each.
- Which rules to follow: we work on copies, never on originals; you mark data gaps as MISSING instead of guessing; you ask before overwriting anything.
That third section makes the biggest difference, because it turns the agent's general good intentions into specific, checkable behaviors.
Step 4: A first warm-up task
The first task should be cheap, fast and risk-free. Take your product catalog export, make a copy in the working folder and ask the agent for a data quality report: "Read the file catalog.csv. Produce a report: missing fields, duplicates, inconsistent units and formats. Change nothing, only report." That is an hour of work including reading the output. The agent modifies nothing, so there is nothing to break.
And the payoff is double:
- Value on day one: a report like this almost always finds problems nobody knew about.
- Learning to delegate: you learn how to phrase an instruction, what you forgot to specify and what a good answer looks like. That skill pays off at every step that follows.
Step 5: A first workflow with a gate
Now the first real workflow, in draft-first mode: the agent drafts, a human approves. A good candidate is descriptions for twenty products, written to the rules in brand/ and products/. Twenty, not two hundred: the point is learning the loop, not volume. You review every draft and make corrections.
And here is the heart of the whole framework: corrections do not get lost. Every note that comes up a second or third time gets written into the knowledge hub:
- The agent over-promises? A rule about forbidden claims goes into brand/.
- It keeps mixing up dimension formats? A pattern with an example goes into products/.
Corrections go back into the knowledge. This one loop separates a system that works better every week from a toy that keeps making the same mistakes forever. A review without writing the rule down is work thrown away after one use. A review that ends with a written rule is an investment every future task draws on. That is how compounding works here.
Step 6: Rhythm and growth
The system needs a rhythm, otherwise it dies after two weeks of enthusiasm. A weekly half-hour review with three questions is enough:
- What worked this week?
- What did we correct more than once?
- What do we therefore add to the knowledge hub?
Only when the first workflow is stable do you add the next ones, one at a time: catalog translations, feeds for more channels, a weekly sales report. You do not have to write everything from scratch either: in our free AI Library we publish ready-made prompts and an open set of skills for agents that you plug in instead of inventing your own instructions from zero.
There is also a boundary worth knowing. As long as the agent works on copies in the sandbox, you can experiment on your own, safely. When a workflow is meant to touch production, integrate with your ERP or process customer data, talk to a technology partner before switching it on. That is a different class of risk and a different level of responsibility.
Everything else, meaning the six steps in this chapter, is within your reach starting today. A folder, a copy of the catalog, a first instruction. The rest is rhythm.
Questions
Do I need to be technical to roll out this framework?
No. The only moment that smells of tech is installing Claude Code, and you do that once, following the guide. All the work after that is writing instructions in plain language and keeping folders in order, which are skills you already use daily. The hardest part of the framework is not technical at all: it is writing down your brand, data and process rules, and no IT person can do that for you.
How much does it cost?
Claude Code requires a paid subscription or usage-based fees, and plans and limits change, so check current pricing with the provider before you start. Beyond that, the framework has no costs: a folder and plain text files are free, and so are the ready-made prompts and skills in our AI Library. The biggest investment is the time it takes to write down your company knowledge. You pay that cost once, and every future task draws on it.
What if I have very little time?
Do steps 1 through 4 and stay there for a week or two. That is realistically two evenings of work, and the data quality report from step 4 delivers value immediately with zero risk. The knowledge hub does not have to be complete at the start: you begin with a few files and add rules as corrections come up, instead of blocking out a week for one big writing session. In this framework, order matters and pace does not.
