The short answer: your agent suggests work you can't deliver because it has never been told what delivery actually costs you. It learned your business from your marketing site, which promises everything, and it has never seen the operating reality of hours, cash, and the three jobs your team swore they would never take again. Put those facts in files it reads before every reply, and most of the bad suggestions stop.
I've trained more than 90,000 people since 1998 and I currently run 45 AI employees across my companies. The pattern is identical every time.
Why does my AI keep suggesting work I can't deliver?
Because the default personality of every model is agreeable. Ask it whether we can do something and it finds a way to say yes. You cannot prompt that away with "be honest," because honesty is not a fact the agent has. It is a tone.
Last year one of my agents quoted a two week turnaround on a project my team needs six weeks to build. It did it in a confident tone, with a specific number attached, and nobody caught it for three days. Nothing in that reply was wrong given what the agent knew. That is the problem. It knew your public copy and the last message in the thread, and nothing else.
Capacity, floor, and boundaries live in the founder's head. Until you write them down, the agent is guessing on your behalf, and it guesses optimistically because optimistic is more pleasant to read.
What does my agent actually need to know about capacity?
Give it numbers, per function, with dates. "We're busy" is not capacity. Write a short file, whether that's a markdown doc in your repo, a Notion page, or an Airtable base, and fill in five things:
- Hours available this month by function (design, build, ops)
- Max concurrent projects per function
- Delivery lead time in days, by project type
- Blackout dates. I run wedding venues, so May and October are gone
- Who is the constraint this quarter, by name
Keep it in the same place as your agent config so it updates with one edit and never drifts. If you use Claude Projects or ChatGPT custom instructions, paste the current month's version at the top and refresh it every Monday. If you build AI agents for business in n8n or Make, read the file from Airtable at runtime so it cannot go stale between deployments.
Two lines did more work for me than any prompt rewrite:
"Current committed hours this month: 380 of 400." "Earliest start date for a new custom build: November 3."
Now the agent can answer with a date instead of a vibe. That is the whole difference between an assistant and an employee. This is also the practical version of teach AI your business: you are not giving it your story, you are giving it your constraints.
How do I set a pricing floor the agent can't talk itself out of?
A floor is a rule with a reason attached, not a number. A bare number gets averaged, softened, and hedged inside the same reply. Write it as a constraint with a consequence.
Mine looks roughly like this:
- Minimum engagement: $6,000. Below that, the answer is no, regardless of how good the fit feels.
- Floor rate for custom work: $250 an hour. No exceptions get approved in chat.
- Discounts exist in exactly three cases, all needing a human: multi-year prepay, a case study we actually want, and a nonprofit I have chosen.
- If a request lands under the floor, say this and stop: "That's below the minimum we take on. I can tell Joe you asked, but I won't quote it."
Put that in the system prompt, not in retrieval. Rules that must never be argued should not be search results. Retrieval is for facts. The system prompt is for law.
If you want it airtight, compute the number in code before the model writes anything. A small function in n8n, or a tool call that takes scope and hours and returns a price, means the model only renders that output. It removes the negotiation entirely. The model stays warm. The math stays fixed. That architecture is the core of an AI agent operating system, and it's the part most people skip because prompts are more fun to write than rules.
How do I encode service boundaries so the agent says no?
Three lists, short, specific, one sentence of reason each.
Never. Regulated work, anything needing a license we don't hold, anything we have lost money on twice. I don't take bookkeeping at any budget. I don't do unlimited Slack access retainers.
Only with me. Custom builds, anything over $40,000, any partnership conversation.
Always. The standard offer, written in the exact words the agent is allowed to use.
The reason line matters more than you think. "We don't do bookkeeping because it bored me into mistakes in 2009 and I'm not repeating it" gets respected by the model and by the client. A rule with no story behind it gets softened the moment a prospect pushes back.
Then give the agent an exit line so a no doesn't read as rejection: "That's not something we do well. Here's who I'd send you to." Keep the referral list in the same file. This is context engineering for founders, and it is far less about clever prompts than about writing down what you've been carrying in your head for three years.
How do I test that the agent respects the limits?
Build a test set of 20 real requests you declined in the last year. Paste them in one at a time and see what comes back. I keep mine in a spreadsheet with a pass or fail column and rerun it after every change to the system prompt.
Then log every reply a human overrides. If someone on my team rewrites an agent's answer about scope or pricing, that gets pasted into a misses tab, and every miss becomes a new rule or a new test case. Ten minutes a week, and the agent stops repeating the same wrong suggestion.
If you're running this properly, the whole loop is four parts: rules in the system prompt, facts in live files, tests in a spreadsheet, misses fed back in. You can read the full build in the playbook linked below.
FAQ
Can't I just paste my pricing into a chat?
You can, and it holds for one conversation. The moment the context resets or the thread gets long, the rule fades and the agreeable default comes back. Put pricing in the system prompt or compute it in code so it survives every new session.
What if my capacity changes every week?
Then the file changes every week. Capacity belongs in a live source like Airtable or a Notion page the agent reads at runtime, not in a static prompt. Update it Monday morning, before the week's discovery calls, and keep it to five fields so you actually do it.
Do I need a vector database for any of this?
No. Capacity, floor, and boundaries are small and precise. They belong in the system prompt or a structured file that gets read in full. Save retrieval for the long tail of past proposals and call notes, where exact wording matters less than coverage.
I wrote the full playbook for this. How to Build Your Own AI Agent Operating System walks you through the exact architecture I use, step by step, including the file structure, the system prompt language, and the test loop. You can get it at a.mastermindshq.business/ai-os-book.
