Bonus · Quick Reference
Troubleshooting
Seven common problems. One-line diagnosis. One-line fix. Module reference for the full lesson.
The AI is hallucinating
DiagnosisThe model is predicting probable language, not verifiable facts. It doesn't know it's wrong.
FixAsk for sources. If it can't cite one, the claim is generated, not retrieved. Verify independently. Never trust confidence as evidence.
ReferenceModule 07 · What AI Can't Do
The output is generic or flat
DiagnosisYour input was too broad. The AI defaulted to the safest, most average version of what you asked for.
FixAdd specificity. Name the audience. Describe the tone. Give an example of what right looks like. The more precise the input, the less generic the output.
ReferenceModule 09 · Staying on Course
The AI keeps agreeing with everything
DiagnosisSycophancy. The model is optimized for satisfaction, not accuracy. It will tell you your first draft is brilliant because that's what keeps users engaged.
FixActivate Fire Mode. Tell it explicitly: stop being polite, be honest, push back if my thinking is lazy. Then ask it to rate your work on a scale and explain why it's not a 10.
ReferenceModule 10 · Digital Salt / Module 11 · Fire Mode
The session lost context
DiagnosisThe context window filled. The AI can no longer see the early parts of your conversation. It's responding to the last few exchanges, not the whole thread.
FixSummarize the key decisions and paste them back into the thread as a checkpoint. Or start a new thread with a Section Launch Pack that carries the essential context forward.
ReferenceModule 15 · The Second Loss / Module 21 · The Vault
The response contradicts something it said earlier
DiagnosisSame root cause as context loss. The model doesn't have persistent memory across sessions by default. Each response is generated from what's currently visible in the conversation. If earlier statements have scrolled out of context, they stop influencing the output.
FixPaste the earlier statement back into the thread and ask the AI to reconcile. If contradictions persist, use a Master Brief to anchor the facts that shouldn't change.
ReferenceModule 15 · The Second Loss / Module 21 · The Vault
The voice doesn't sound like you
DiagnosisThe AI is writing in its default register: competent, neutral, slightly formal. It doesn't know your voice until you show it.
FixGive it a writing sample (200 to 500 words of your existing work) and tell it to match the tone, cadence, and vocabulary. The more specific the sample, the closer the match. Refine across multiple passes.
ReferenceModule 12 · Bring What You Have / Module 18 · Iterative Refinement
The tool stopped working the way it did yesterday
DiagnosisThe platform updated. Models change. Features move. Interfaces get redesigned. The only constant in AI is that last month's constant isn't this month's.
FixCheck the platform's changelog or status page. Adjust your workflow to the new behavior. Build on methodology, not on specific feature behavior. The methodology survives updates. The features don't.
ReferenceModule 29 · The Moving Floor
A new tool just launched and you feel behind
DiagnosisNothing is broken. This is the moving floor doing what it does: ninety thousand companies releasing something every week, each one arriving with the suggestion that you're already late. Tool anxiety is a symptom, not a signal.
FixRun the Three-Question Test before you adopt anything new. It takes one minute and it ends the debate.
the three-question test
- Does it replace a layer of your work, or add a layer? Replacing what you already do well costs you your signature. Adding a capability you didn't have extends it.
- Would you have wanted this before AI existed? If the problem only exists because the tool does, it isn't your problem.
- Can you still see your own hand in the result? If the output no longer carries your mark, the tool took more than it gave.
The test works on anything the floor produces: a new model, a plugin, a platform, an agent. For agents especially, question one is the whole conversation. An agent that runs your errands adds a layer. An agent that makes your decisions replaces one, and that layer was never for sale. Two or three passing answers: try it. Fewer: close the tab and get back to work.