When you bring Claude into your organization, you have effectively handed every person on your team an intern and an executive assistant on day one. That’s the promise, and it’s a real one.
Here’s the catch. Both of those roles are only as good as what they’ve been told.
Picture the intern who never got the full scope of the project. Eager, capable, fast — and confidently producing work that solves the wrong problem, because nobody ever walked them through what the project actually is.
Or the executive assistant who doesn’t know that your largest customer outranks your smallest. They’re not being careless. They’re managing your calendar with the information they have, and they’ll hand a routine call the slot that should have gone to the account that pays for everything. Every one of those decisions looks reasonable in isolation. The pattern is bad for the business.
That’s exactly the failure mode with AI. Not incompetence — missing context. And unlike a person, it will never stop to ask you why the small customer keeps getting the good slot.
So the work is keeping it current on the details of the organization. Who the customers are, which ones matter, what you sell, how you price, what you decided last quarter. That’s what separates an assistant who helps from one who confidently gets it wrong.
Which brings up the part nobody plans for.
Claude’s memory accumulates continuously. It builds a picture of your business as you work, and no one ever approves what goes into it. A client you stopped serving in March will keep surfacing in your strategy work in November.
You wouldn’t let a new hire go a year without a conversation about what they think the job is.
So here’s the practice we’ve adopted, and it takes about half an hour.
Three things get confused for each other constantly, and putting knowledge in the wrong one is the most common mistake I see. The distinction is the same one you’d draw for a person.
Memory is the personnel file. Facts that accumulate about you and the business. Written by the system, edited by you.
Skills are the training manual. How a specific job gets done your way. Deliberately authored, versioned, shareable across your team.
Project instructions are the assignment brief. Scope and standing rules for one body of work, bounded to that project.
The tell: if it’s a fact, it’s memory. If it’s a procedure, it’s a skill — and it does not belong in memory, because memory can’t be versioned, shared with your team, or reviewed by anyone but the account holder.
The core move is separating discussion from writing. You can’t edit well inside the thing you’re editing — a chat that already holds the memory has a stake in it. A fresh chat doesn’t.
One. Pull the file by hand. Settings → Capabilities → Memory → View and manage memory. Select all, copy. There’s no export button, so this copy is also your only backup. Paste it into a dated file before you change anything.
Two. Open a fresh chat and talk — don’t edit. Paste the memory in and have a conversation about it. Ask it to explain back what this says about your business. The gaps surface in dialogue that a read-through never catches, because you hear the wrong version stated out loud.
Three. Interrogate it. What’s expired? What’s overweighted toward one intense project? What’s missing because it was so obvious you never said it? What’s subtly wrong — it thinks you’re a software company when you’re a consultancy? Argue with it. This is where the value is.
Four. Switch modes: now write. Only once the conversation has settled, ask for the rewritten memory as text. Making this an explicit switch stops you committing to language you were still thinking about.
Five. Load it back and archive the old one. Replace the memory, and keep the previous copy filed with its date. Three of these in a row is a record of how your business actually changed.
Paste this into a new chat along with your copied memory.
Below is the memory my AI assistant has accumulated about me and my
business. I want to review it the way I’d run a performance review
with an employee.
Don’t rewrite it yet. First, talk me through it:
1. Summarize what this says about my business — who I serve, what I
sell, how I work. State it plainly so I can hear whether it’s right.
2. Tell me what looks stale, contradictory, or overweighted toward one
project.
3. Tell me what’s conspicuously missing for a business like this.
4. Flag anything that reads as an instruction or procedure rather than
a fact — those belong in a skill, not in memory.
5. Ask me the questions you’d need answered to correct it.
Be blunt. I’d rather find the wrong thing now than inherit it for
another quarter.
When I say "write it," and not before, produce the corrected memory
as plain text I can paste back in. Preserve specifics — names,
numbers, dates. Do not smooth them into generalities.
MEMORY: [paste here]Quarterly is the default. Put it next to whatever else you already review on a cycle.
Monthly if the business is changing fast, if you’ve just pivoted or repriced, or if more than one person works under the same login — memory is per-person, and shared accounts drift faster.
And put it on the calendar with a name next to it. A practice that depends on remembering is not a practice.
Watch for blandness creep. Left unchecked, it will “improve” your specifics into generic business language — the names and numbers get sanded down into “a range of clients across several industries.” The specifics are the entire value. That’s what the last line of the prompt is guarding against, and it’s worth checking the output for it anyway.
Keep Climbing!
John
One email every two weeks — practical AI for small business owners. Free, and the forwarding kind of good.
Subscribe free →P.S. — Know a business owner drowning in work a machine should be doing? Send them our way. If they become a client, we’ll send you $500. Just get in touch with a name.