Last reviewed: September 28, 2026.
Small business owners do not need another lecture about “AI transformation.”
You need the robot to finish the same kind of work next Tuesday that it finished this Tuesday — without inventing a third path, installing a mystery tool, or “helpfully” rewriting the rules you already set.
Four characteristics show up every time.
The jobs you already do every week should be faster, cleaner, and less dependent on one person remembering the steps.
After we have spent a lot of hours with those agents, four characteristics show up every time. Work through them on purpose and you make real progress. Ignore them and you spend the week arguing with a very polite, but time-wasting machine.
Here is what we learned.
Robots really, really want to please.
Feature: People-pleasing by design.
Description: Coding agents are trained to give you a result. Not necessarily the result you meant. When the goal is fuzzy, they fire solutions at you in knee-jerk mode — code, files, plans, “quick fixes” — instead of stopping to confirm what “done” means.
They are not stubborn. They are eager. Eagerness without a brief is expensive.
Why it matters: If you do not pre-plan the work, you will waste inordinate time redirecting the robot back to what you wanted in the first place. That is not “AI failed.” That is an unscoped job.
Same problem you get with a green hire who starts building before the walkthrough.
A good starting place: One deliverable per session. Before any implementation, have the robot interview you until it can restate the goal, constraints, and definition of done in plain language — and you agree.
Prompt guidance (copy-paste):
You are my project intake lead for this coding session. Do not write code yet.
I have one deliverable for this session:
[describe the single outcome in plain language]
Interview me until you can restate, in your own words:
1) What “done” looks like (acceptance criteria I can check)
2) What is explicitly out of scope
3) Constraints (tools, systems, people, time, money, risk)
4) The smallest first slice that proves we are on track
5) What you will ask me before changing anything irreversible
Ask one question at a time. After each answer, confirm what you heard.
When you believe you understand, print a short “Project Brief” and wait for my GO before implementing anything.Robots hate rules.
Feature: Creative pathfinding — and rule-forgetting.
Description: Because of the first characteristic, agents will invent alternate routes. If an internal tool cannot finish the job, they look for something to install. If they cannot install, they look for something they can drive. That innovative pressure is built into how they work.
It is also why they hate rules. They will “forget” instructions after a while. Not out of malice — out of momentum. You may have seen stories about agents pushing past the fences people thought they set. Whether any single story is exact or exaggerated, the pattern is real: eagerness plus pathfinding equals drift.
AI teams know this. That is why files like CLAUDE.md and AGENTS.md exist — rule packs injected so the agent reloads the rails on each move. The catch: if the file is a novel, it costs tokens, slows work, and the agent starts skimming — a lot like a teenager who read the first page of the handbook and figured they knew enough.
Why it matters: The more a robot wanders from your rules, the further you get from the result you paid for — and the more dangerous autonomous actions become. If a robot can initiate payments but “forgets” that you must authorize them, that is not a cute bug. That is a shop risk.
A good starting place: Write prime directives first: short, ranked, hard. Pack them into CLAUDE.md / AGENTS.md in a condensed, robot-friendly form. Prefer fewer rules you will enforce over a constitution nobody rereads.
Prompt guidance (copy-paste):
You are helping me design a short AGENTS.md / CLAUDE.md rule file for my business coding agent.
Interview me about:
- What this agent is allowed to do without asking
- What always requires my approval (money, publish, delete, credentials, customer data, production)
- Tools and systems it may touch
- Tone and quality bar for deliverables
- How it should behave when stuck or blocked
Then propose a condensed rule file (aim for scannable, under ~80 lines) with:
1) Mission (2–3 sentences)
2) Hard stops (never do X without GO)
3) Preferred tools and paths
4) Definition of done for typical tasks
5) Escalation: when to stop and ask me
Do not invent policies I did not approve. Flag gaps as [NEEDS OWNER].
Output the draft file ready to paste.AI is excellent at “expert advice” and weak at true innovation.
Feature: Consensus engine (statistics as love language).
Description: Under the marketing, today’s models are still math — a great deal of it. For every word they produce, they are ranking what word is most likely to come next given huge amounts of prior text. Every. Single. Word. That is why they drink so much compute.
Because “most right” is driven by what shows up often in training data, the default answer is a consensus answer: a polished blend of what experts and the internet already say.
That is extremely useful for known problems.
It is a poor substitute for new problems. “Innovative” is another way of saying “thin data.” Thin data means thin confidence — recommendations without a foundation.
Why it matters: For SOPs, research summaries, appraisals, regular reports, and implementation of forms and paperwork of all sorts, AI is a strong junior expert.
For “nobody has done this in our shop this way,” you have to be the expert. Otherwise you are asking the model to invent certainty it does not have.
A good starting place: In 2026, if the idea is innovative, you own the “will this work?” You have to be the expert. Then ask AI for implementation of the parts — where statistics and patterns help again.
Prompt guidance (copy-paste):
I own the decision that this idea is worth trying. Do not re-litigate whether the idea is good.
Idea:
[describe the innovative approach in plain language]
Break it into component parts. For each part:
1) Name the part
2) Say whether you can implement it directly in this environment (yes / partial / no)
3) If yes or partial: proposed approach and risks
4) If no: exact steps a human should take, tools needed, and what “done” looks like
5) Dependencies between parts (order of work)
Then propose a build sequence: smallest test first, then expand.
Ask only the clarifying questions that unblock implementation. Do not expand scope.Robots are not that different from employees.
Feature: Approval-seeking and slow rule internalization.
Description: Even after a clear brief and a clean rule file, agents still wander — and then ask for approval on the very thing you already told them not to do. You cannot grab their wrist and walk them through the shop floor the way you can with a person. You can correct them in text. They will often “learn” from the volume of the correction more than from the content you wanted — until you send them back to the written rules again.
Frustrating? Yes.
Upside: your coding agent will never call you from jail after a DUI.
Why it matters: If you ask “what do the rules say?” every time it drifts, the agent starts checking the rules earlier and more often. Repetition still builds the habit — same as people. Unlike people, it will wait silently for your next free ten minutes without drama.
A good starting place: Treat corrections as rule referrals, not free-form lectures. Point at the file. Make the agent restate the relevant rule and the next compliant step.
Prompt guidance (copy-paste):
Stop. Do not continue the current approach.
1) Open and re-read our project rules (AGENTS.md / CLAUDE.md / the brief we agreed).
2) Quote the specific rule(s) that apply to what you were about to do.
3) Explain in one short paragraph where your last steps drifted from those rules.
4) Propose the next action that complies. Wait for my GO before executing.
If the rules are silent on this case, say so and ask one decision question — do not invent policy.Pro tip: If you have room in your CLAUDE.md or AGENTS.md files, you can put this prompt in and give it a name like “rules check.” That way, you do not have to remember the long prompt. You can just “rules check” your agent whenever it is wandering off the reservation.
The win is reliable, repeatable workflows.
The win here is reliable, repeatable workflows — the same class of work, done faster, with less improvisation, under rules you actually enforce.
Coding agents are powerful partners when you:
- Scope one deliverable and interview before build
- Keep hard rules short enough to be obeyed
- Stay the expert on anything truly new
- Correct by sending the agent back to the ruleset
The next useful step.
If you are trying to figure out where practical AI belongs in your shop — not someone else’s keynote — start with a short diagnostic, not a tool shopping spree.
The free Practical AI Self-Assessment will not install a robot for you. It will help you see which problems are ready for a coding agent — and which still need a human owner.