What to NOT do
when you use AI.

::symptom

Vague, generic output that you have to re-prompt three times to make useful.

::fix

Use the 7-part prompt: Role · Context · Input · Task · Shape · Audience · Voice. State each one explicitly.

::symptom

First draft is wrong-shaped. You waste tokens correcting structure when you should have caught it at outline.

::fix

'Don't write yet. Outline first. I'll approve, then you write.' This single move saves more time than any other prompt habit.

::symptom

Three variants of the same idea. The 'options' are all the same shape.

::fix

Ask for 12 variants across 4 explicit registers. The constraint forces real range.

::symptom

Average-quality output that you ship because it's there.

::fix

Pass 2 and 3 are where the unlock happens. 'Now give me a contrarian version' and 'now the version that would surprise the smartest person in the room.'

::symptom

Generic, defensive, on-the-fence answers.

::fix

'You are a senior X who has done this 100 times.' Roles unlock specificity.

::symptom

Output is the wrong register for who reads it.

::fix

State the audience explicitly. 'A skeptical board.' 'A 12-year-old who has never seen this.' 'A grad student in this specific field.'

::symptom

Diplomatic, hedged, soft output. AI defaults to flattery.

::fix

'Be direct. Skip the diplomatic version. Tell me what would embarrass me at [peer audience].'

::symptom

AI focuses on irrelevant parts. Token cost balloons. Reasoning quality drops.

::fix

Paste only what's needed for the specific task. If you need long-context analysis, use the model built for it (Claude 200K, Gemini 1M).

::symptom

You walk away with the answer to the question you knew to ask. You miss the question you didn't.

::fix

Always end with 'what should I have asked that I didn't?' or 'what's the question this conversation is dodging?'

::symptom

AI hallucinates a paper that doesn't exist, or paraphrases an existing paper incorrectly.

::fix

Pull every citation manually. Google Scholar. PubMed. Westlaw. Whatever the authority is for your domain. AI is the researcher; the source is the truth.

::symptom

Arithmetic errors in numeric output that pass the eye test. Decisions made on bad math.

::fix

Verify in Wolfram, Python, or a textbook. Every number that matters. AI is bad at math; this is real, not a 2023 leftover.

::symptom

Acting on advice that's structurally wrong because it was tuned to be helpful, not licensed.

::fix

AI is the patient-advocate friend, not the doctor. Use it to prep questions for the real professional. The professional gets the call.

::symptom

Voice dies. Audience notices. Trust never recovers.

::fix

AI is the draft. You ship the draft after a human pass. Always. Especially in voice-driven channels (newsletters, social, YouTube).

::symptom

Every idea you bring gets validated. You stop being challenged.

::fix

Force the steelman. 'Now argue against this from the perspective of someone who would lose money if it works.'

::symptom

AI's most-wrong outputs are often its most-confident.

::fix

Confidence is a stylistic feature of the output, not evidence of accuracy. Always verify when the cost of being wrong is real.

::symptom

Patient data, customer SSNs, internal financials sitting on a vendor's servers you can't audit.

::fix

Use local Ollama for sensitive work. Or your facility's vetted BAA-covered internal LLM. NEVER ChatGPT.com / Claude.ai for protected data.

::symptom

Credentials leak. Even 'private' chats are subject to subpoena, breach, vendor employee access.

::fix

Treat every cloud AI conversation as if a stranger might read it tomorrow. Because they might.

::symptom

Rm -rf, force-push, prod-DB-truncate. The model didn't mean to. You ran it anyway.

::fix

Read every command. Especially the destructive ones. Especially when you're tired. The AI doesn't carry the consequences.

::symptom

Inappropriate content surfaced. Personal info shared. AI-companionship-for-isolation patterns.

::fix

Age-appropriate conversation about what AI is and isn't. Adult co-using for under ~14. Talk about what to never share.

::symptom

AI gives you scripts and frameworks. Real distress doesn't move. Sometimes worsens.

::fix

Real therapy is real human contact, calibration, accountability. AI is the journal between sessions, not the therapist.

::symptom

Lots of opinions. Zero actual prompts run. The skill never compounds.

::fix

Type into the chat box. Six times. Today. The skill is in the repetition, not the reading.

::symptom

Result is close-but-wrong. You give up instead of iterating.

::fix

Most real work is 3-5 turns. Start with outline. Then draft. Then critique. Then rewrite. Then ship.

::symptom

Every new AI session starts from scratch. You re-teach the AI your context every time.

::fix

Save your best prompts as reusable files. Load them at the start of every session. Over weeks, this is the single highest-leverage skill.

::symptom

You miss what other models are good at. You hit one model's blind spots repeatedly.

::fix

Different models for different tasks. Use the decision tree. Cross-check important work in two models.

::symptom

You're slower than the AI is fast. The friction kills the habit.

::fix

Set up the AI in your text editor / OS / phone with shortcut keys. The friction-to-prompt should be under 2 seconds.

::symptom

ChatGPT + Claude + Cursor + Perplexity = $80+/month. You use 2 of them seriously.

::fix

Pick one paid subscription (probably Claude OR ChatGPT). Use the free tier of everything else. BYOK API for power-users only.

::symptom

Notion AI, Google Gemini-in-Workspace, GitHub Copilot... paying for what your existing tools already include.

::fix

Audit your subscriptions. Many give you AI for free now. Pay only for what you'd notice missing.

::symptom

Daily-cap hit. Frustration. You start paying for something you didn't need to.

::fix

Local Ollama (free, your hardware) handles 80% of daily-use cases. Reserve cloud calls for the work that genuinely needs frontier quality.

::symptom

You stop hiring because 'AI can do it.' Then the work hits a quality ceiling you can't break through.

::fix

AI is leverage, not replacement. The human who works WITH AI is the strongest configuration. Don't fire the human; arm them.

::symptom

Your roadmap, your pricing, your positioning, all come from AI conversations. You become a feed-forward network with no taste.

::fix

AI is the thinking-partner, not the strategy oracle. The decisions that distinguish your business are the ones you make. AI critiques, you decide.