When AI gives you something useless, the tool is rarely the problem. One of five layers is missing, and the symptom tells you which one. This is the diagnostic I use in the workshop, because a business owner who can name the missing layer fixes their own output instead of concluding that AI does not work for their industry.
Key takeaways
- Every common failure maps to a missing Context Stack layer: Context, Task, Source, Rules or Output.
- Generic writing is a Context problem. Invented facts are a Source problem.
- Gap-filling is a Rules problem, and it is the one that gets businesses into trouble.
- Fix one layer at a time and rerun the same example, or you learn nothing.
- Output that was good once and useless the next time means there is no Stack, only a conversation.
The diagnostic table
| What you are seeing | Missing layer | The fix |
|---|---|---|
| Reads like a brochure, could be any business | Context | Two sentences on who you are and who is reading |
| Answered a different question | Task | One sentence naming the exact document you want |
| Invented a price, figure, date or detail | Source | Supply the actual price list, notes or approved answers |
| Filled a gap instead of asking | Rules | Add the rule: flag anything not covered, never assume |
| Right facts, unusable shape | Output | Describe the sections, order and what stays separate |
| Confidently wrong about your trade | Source, then Rules | Real documents first, then limits |
| Too long and padded | Output | State the length and what to leave out |
| Too vague to act on | Task and Output | Name the document and its required contents |
| Good yesterday, useless today | No Stack | Write it down instead of retyping requests |
| Sounds like everyone else's marketing | Context and Source | Your own words and your own material |
| Gave advice it should not give | Rules | The escalation rule, written before the answering rule |
| Different answers from different staff | No shared Stack | One agreed Stack per job |
The three failures worth understanding properly
Invented detail. This is the failure people describe as the AI lying. It is not lying, it is answering with nothing to answer from. An assistant given no price list will produce a price, because you asked for a quote. The fix is never a cleverer instruction. It is attaching the file.
Gap filling. More dangerous than invention, because it looks correct. The assistant meets a hole in your source material and smooths over it. The fix is one line in the Rules layer: if information is missing, list it at the top and do not proceed on an assumption. Test it by deliberately removing a measurement and confirming it complains.
Format drift. The quiet killer. The content is right, the shape is wrong, so somebody reformats by hand every time, and within three weeks the workflow is abandoned because it is not actually saving anything. The fix is a specific Output description, not "make it professional".
How to debug properly
One. Keep the example fixed. Use the same job every time so you can compare outputs.
Two. Change one layer. If you change three and it improves, you have learned nothing about which one mattered.
Three. Rerun and compare against a document you already sent. Your own past work is the benchmark, not your impression of the draft.
Four. Write the fix into the Stack. A fix that lives only in your head will be missing next week.
Five. Keep a short list of what broke and what fixed it. That list becomes your Rules layer, and it is the most valuable page in the whole system.
What is not a layer problem
Some work genuinely should not be handed over. A pricing decision, a clinical judgement, a repair method, anything a licence covers, anything where being wrong puts a person or a house at risk. If the output is wrong because the job needs human judgement, no amount of layer fixing helps. The correct answer is to have the assistant prepare the document and a person make the call.
Read the full Context Stack, or learn how to check output before it goes out.
Fix yours in the room
Bring the output that is not working and the job behind it. A full day at Avani Mooloolaba Beach Hotel on the Sunshine Coast, A$495 excluding GST, capped at 15 people, and you leave with the Stack rebuilt and tested against your own example.
See the workshop, or enquire about onsite team training.
Frequently asked questions
Why does AI make things up?
Because you asked it for something and gave it nothing to work from. It produces the most plausible version of what you requested. Supply the real material and the behaviour changes immediately. That is a Source layer fix.
How do you stop AI inventing prices?
Attach the price list as source material and write the rule that says use only supplied prices and flag anything missing. Then check figures against the list before anything goes out.
Why does AI output sound generic?
Missing Context and Source. With no description of your business and none of your own material, the only thing left to write from is the average of everything, which is what generic means.
Why was it good once and bad the next time?
Because it was a conversation, not a written Stack. Different wording each time produces different results. Writing the layers down removes that variance.
Is a longer prompt better?
No. A complete one is better. Length without your source material and your rules just produces more confident generic writing.
How do you test whether the rules are working?
Deliberately break the input. Remove a measurement or a price and confirm it flags the gap rather than filling it. If it proceeds anyway, the rule is not strong enough.
Should you switch AI tools if the output is bad?
Rarely. The same missing layers produce the same failures in every tool. Fix the Stack first, then compare tools on the same complete input. Compare ChatGPT and Claude for business.
Why do staff get different results from the same tool?
Because each person is supplying different context and rules. One agreed Stack per job fixes it. How to set team rules.
What if the output is right but nobody uses the workflow?
Format drift is the usual cause. The shape is wrong, someone reformats by hand, and the saving disappears. Tighten the Output layer. How to measure whether it saved time.
How long should debugging take?
Minutes per cycle once the example is fixed and you change one layer at a time. Most workflows stabilise after three or four rounds.