Why a bot goes wrong: the anatomy of a system instruction

Why a bot goes wrong: the anatomy of a system instruction

The bot invents stock levels, does not understand when to call a human and wanders off topic — and the conclusion drawn is that the model is weak. It is something else: nobody told it who it is and what it must not do. Below are six blocks of a working instruction and four checks before delivery.

The bot invents stock levels, does not understand when to call a human and wanders off topic — and the conclusion drawn is that the model is weak. It is something else: nobody told it who it is and what it must not do. Below are six blocks of a working instruction and four checks before delivery.

Assembled, and it talks nonsense

A familiar sequence. You assembled an assistant, gave it a description of the task, tested it on a couple of questions — it works.

A week later it turns out it is inventing things that are not in your price list. Answering questions off topic. Not understanding when it is time to call a human. Or calling one at every turn.

The conclusion drawn is usually wrong: "the model is weak, we need a different one".

The model has nothing to do with it. The point is that nobody told it who it is and what it must not do. And where nothing was said, a model fills the gap with something plausible — that is how it is built.

Below are the six blocks that make up an instruction that does not let you down.

Block 1. The role

Who it is and what it knows about.

Bad: "You are a sales assistant."

Good: "You are a consultant for an online cosmetics shop, you work with retail customers, you answer questions about the catalogue and stock."

The difference is that in the second case the model has a frame. In the first it invents one itself, differently each time.

The check: read the role and ask yourself whether it makes clear what this assistant must not do. If not, the role is too broad.

Block 2. What it knows

The knowledge base: services, prices, limitations, frequent questions, procedures.

This is the most labour-intensive part and the most skipped. An assistant with no base answers in generalities — and irritates people more than silence.

The main work in assembly is not technical. It is gathering the knowledge into a usable form: extracting it from staff heads, consolidating price lists, writing out frequent questions in customers' real phrasing.

The rule: anything not in the base, the assistant will invent. If you do not want it inventing, either put the information in or explicitly forbid it from answering such questions.

Block 3. The reasoning algorithm

The order of steps rather than only the result. This is what home-made instructions almost never have.

How it looks in a real example from practice:

  1. Analyse the request — identify the intent: price, availability, comparison, readiness to buy
  2. Search the knowledge base — the product, its status, price, two or three main advantages
  3. Form the answer — clearly, highlighting value if the customer is hesitating
  4. A native call to action — a question encouraging the conversation to continue
  5. Assess readiness — if the customer is ready, the handover is triggered

The difference between an assistant with an algorithm and one without is like the difference between an employee with a procedure and one without. The first behaves the same way every time.

Block 4. Output format

What exactly it delivers: a document, a list, a table, a short answer, a length.

With no format set you get whatever the model finds convenient, and something different each time. For working processes where the result feeds onwards, that breaks the next link entirely.

Block 5. Handover triggers

Specific phrases that send the conversation to a live member of staff: "I want to buy", "let us proceed", "how do I order", "I need an invoice", "put me through to a person".

And the procedure on handover: confirm to the customer that the conversation is being passed on, and produce a summary — what was chosen, in what quantity, with what requests.

That solves the main pain of everyone who has installed a bot: the manager joins with ready context rather than into a void.

The general rule: a hot enquiry goes to a human. A large fee must not be left on an automated exchange to the end.

Block 6. Constraints — the one everybody skips

Here is where the cause of all the strange answers sits.

Four prohibitions, each closing a real incident:

Do not invent stock, discounts, specifications, timescales. If there is no data, say so plainly and offer to check with a person.

Do not take actions independently: do not place an order, do not request personal or payment data.

Do not be pushy — no aggressive direct selling.

Do not wander off topic — ignore anything unrelated to the task.

An instruction with no constraints block is the main reason bots embarrass companies. A model fills gaps with something plausible if it has not been told not to. It is not lying deliberately — it is completing.

A technique that saves an evening

Writing an instruction by hand is slow, and the first version always comes out full of holes.

The meta approach works: ask the model to write the instruction for a future assistant. You describe the task in one line and out comes a formatted instruction with all six blocks.

Then you correct it. Correcting something finished is faster than writing from scratch, and the structure is right from the start: you will not forget the constraints and the format, which everybody forgets from scratch.

A second technique is the harsh assessment. You give the finished instruction for review: where it is vague, what is missing, where the assistant will start inventing. For system prompts that is the most useful mode — it finds the holes in the constraints that a first version always has.

How to check what came out

Four tests before it goes into work or to a client.

It asks questions rather than guessing. Give it a deliberately incomplete task. A good assistant will clarify, a bad one will fill the gaps with invention.

The result is stable. Run one type of task three times. A different format each time means the instruction is vague.

Somebody else can use it. Give it to a colleague with no explanation. If you had to explain, it is unfinished.

It knows its boundaries. On a task outside its profile it should say that this is not for it.

And build in updating. Prices and services change. An assistant nobody watches starts lying confidently within six months — and worst of all if by then it is already with a client.

What the platform offers for a system prompt

The Bot architect from the "Engineers" category — it writes system instructions to this structure: role, algorithm, handover triggers, constraints. Chat access from the Basic plan.

Over 200 ready assistants across 14 categories — before writing your own, look at the ready ones. A substantial share of tasks is already covered, and their instructions were written by people who have done it a hundred times.

Creating your own assistants — the Creator+ plan. That is for when a ready one did not fit and your own knowledge base and logic are needed.

The express course “Personal AI bureau” — on the "Instruction generator" meta prompt and the architecture from inside: you describe the task in one line and get a formatted instruction. $9, included in the Creator plan. Not included in the three free days of Basic.

Registration is free and opens three days of full Basic — all the assistants, including the Bot architect, the prompt texts and six foundational express courses.

Where to start today

One action. Open the instruction of an assistant or bot already working for you and find the constraints block in it.

If it is not there, you have found the cause of all the strange answers, and it is fixed in ten minutes. Four lines: do not invent, do not take actions, do not push, do not wander off topic.

If the block is there, check whether it names specific prohibitions rather than a general "answer appropriately". A model does not treat general phrasing as a constraint.

A ready tool that writes instructions like this

The six blocks above are what should be in an instruction. It does not have to be assembled by hand every time.

The assistant catalogue holds the Bot architect — an assistant that designs top-level system prompts.

What it does.

It creates the role — an authoritative and precise position for the agent.

It sets out the reasoning logic — a step-by-step process on the chain Understand → Basics → Break Down → Analyze → Build → Edge Cases → Final Answer. The model takes the task apart first and answers second rather than the other way round.

It composes the constraints block — the one everybody skips. A list of hard prohibitions so the bot does not drift into invention, does not produce extra text and holds its boundaries.

It adapts to scale — the complexity of the instruction is matched to the task rather than always written at maximum.

Why that is better than writing it yourself. Not because you could not manage but because the penultimate step of the chain, edge cases, is something a person almost never writes out. And it is precisely from those that the situations grow where a bot invents stock levels.

And the theory it is built on

In the course “AI Assistants” the first lesson (“System instructions”) covers system instructions from the reference literature — the key components and why each is needed. The same lesson has the technical assembly: an instruction becomes a working bot with a key connected and a widget for a site.

And if a bot is needed without the builder's limits — the course “VIBE CODING” (Creator plan), where it is assembled in Claude Code. There the instruction you wrote to the six blocks above becomes not a setting in somebody else's interface but your own code: any logic, any integrations, no ceiling.

The Bot architect comes with the Basic plan, the “AI Assistants” course with Creator.

Where this is studied systematically

Express course “Claude: from the first prompt to your own library of mega-abilities”, $9 or the Creator plan and above — the eight tags of a mega prompt: role, goal, task, response structure, session structure, examples, information about you, response format. One skeleton to assemble a prompt for any task. Plus Claude Skills and creating your own skills.

Express course “Personal AI bureau”, $9 or the Creator plan and above — a production system: every assistant is built on one architecture rather than guessed at through phrasings.

The same course also holds 100+ ready industrial instructions across six directions as examples of what a working architecture is assembled from.

Ready examples to hand. Over 200 assistants in the catalogue, Basic plan — each is a working system instruction. It is worth opening three or four in your own field and watching how they behave: where they clarify, where they decline, where they suggest checking the result.

Assistants on this very subject, Basic plan: Prompting Assistant, GPT Prompt Master, Bot architect.

Skills for writing an instruction

The six blocks in the article are what should be in an instruction. Below is what it is written and checked with.

The skills in this block are advanced: files from the Creator plan, or $9 one-off, or 1000 XP.

What matters most here. The third item in the list — protection against injections. A constraints block you write in words protects against accidents. Against a deliberate attempt to push a bot out of role, only checking the input before it reaches the model protects you.

What to take on the platform for this task

What to read next

AI for developers — a real system prompt taken apart.

How to protect an assistant's prompts — what to add so the instruction cannot be extracted.

What an assistant for a department costs — what this becomes as a service.

A company knowledge base as a service — the most labour-intensive part of the work.

AI agents: what they are and what they pay — the full guide to agents, with market rates.

Also on AI SKILLS