Prompt Anti-Patterns: How Prompts Go Wrong
Most “the model is dumb” moments are really prompt bugs — vagueness, contradictions, buried instructions. Learn to spot and fix them.
You’ve now got a toolkit — decomposition, chain-of-thought, voting. But the most common reason a prompt fails isn’t a missing advanced technique; it’s a bug in the prompt itself. When the output is wrong, the beginner blames the model and re-rolls; the pro re-reads their own prompt and finds the ambiguity, the contradiction, the instruction buried on line 40. This lesson closes the module by cataloguing the usual ways prompts go wrong. Hold the question: before you blame the model, what should you check in your own words first?
The usual suspects
Most bad outputs trace to a handful of prompt bugs. Vagueness: “summarize this” — how long? for whom? what matters? The model guesses, and its guess isn’t yours (recall specificity from Module 2). Contradiction: “Be thorough but keep it under 20 words” — impossible instructions force it to pick one and seem to “ignore” the other. Buried instructions: the one thing you actually needed is on line 40 of a wall of text, so it gets lost. Over-stuffing: cramming ten tasks into one prompt (the exact case for decomposition, lesson 40). Leading/loaded framing: “Explain why X is the best choice” pre-commits the answer and suppresses honest analysis.
Subtler traps
A few less-obvious ones. Politeness mistaken for instruction: “It would be nice if you could maybe consider mentioning sources” is a hint, not a requirement — if you need sources, say “Include sources.” Negation-only instructions: “Don’t be formal” tells the model what not to do but not what to do instead; positive instructions (“Write casually, like texting a friend”) work better. Assuming shared context: “Fix the usual issues” — the model doesn’t know your “usual.” Format left to chance: if you need JSON or a table, say so and show the shape, or you’ll get prose. Each of these would confuse a competent human handed the same note — which is exactly the test.
Same intent, buggy vs fixed: • Buggy: “Can you look at this and make it better? Keep it short but include everything important.” (vague + contradictory) • Fixed: “Rewrite this paragraph for a busy executive. Cut it to ~3 sentences. Keep only the decision and the deadline; drop background.” (audience, length, and what “important” means — all specified)
The debugging habit: fix the prompt first
The meta-skill is a diagnostic reflex: when output is wrong, assume the prompt is the bug until proven otherwise. Re-read it as if you were a smart stranger with no context — where would you be confused or forced to guess? Fix that, then re-run. Only after the prompt is genuinely clear do you reach for heavier tools (decompose it, add reasoning, vote). This is the same “the operator is part of the system” humility from Module 2, now sharpened: most failures are communication failures, and communication is the half you control. A clear prompt beats a clever one.
Think of briefing a brilliant new assistant who takes every word literally and can’t read your mind. If you say “handle the thing from yesterday, keep it brief but complete, and make it pop,” you’ll get something baffling — not because they’re incompetent, but because your note was vague, contradictory, and assumed they know “the thing.” A great manager doesn’t fire the assistant after that; they re-read their own note, see why it was confusing, and rewrite it clearly. Prompting is exactly that: when the result is off, suspect the briefing before the brain.
Debugging a failing prompt instead of re-rolling: 1. Prompt: “Analyze our sales data and give me insights, be concise but thorough.” Output: generic fluff. 2. Re-read as a stranger: “insights” about what? “concise but thorough” contradicts. No definition of success. 3. Fix: “Here are monthly sales for 2 years. Identify the 3 largest month-over-month drops, state the month and % for each, and give one plausible cause to investigate. Bullet points.” 4. Now it’s specific, non-contradictory, and format-pinned — the ‘dumb’ model produces sharp output. The bug was the prompt.
This is the reading. The interactive version — active-recall quiz, a hands-on experiment you run in your own AI, and an earned mastery check — is free in the app.
Start this lesson free →