Prompt Engineering
The Anti-Patterns (How to Fail)
"Invert, always invert." — Carl Jacobi
Mastering the art of prompting often comes down to understanding exactly what makes a prompt fail. By learning how to write the worst possible prompt, you can guarantee better results by doing the exact opposite.
This guide uses the mental model of Inversion: instead of asking "How do I write a great prompt?", we ask "How do I guarantee the model gives me garbage?"
Here is a systematic list of the most reliable ways to sabotage a prompt, grouped by the "Anti-Pattern" and the Design Pattern that solves it.
The "Lazy Delegator" (Vagueness)
The Mistake: Be as vague as possible. Use broad verbs and ambiguous words like "cool", "better", or "nice".
Why it fails: The model lacks "grounding" (shared physical/social context). Without explicit definitions, it regresses to the mean, giving the most statistically likely (mediocre) answer.
Anti-Prompt: "Write something about marketing."
Correction (Inversion): Define constraints and criteria. "Write a LinkedIn post about B2B marketing trends in 2025. Success criteria: Focus on AI adoption, use a professional but provocative tone, and include 3 bullet points."
Related Strategy: Explicit Intent Protocol — Treat every prompt as a standalone packet containing all necessary definitions.
The "Mind Reader" (Missing Context)
The Mistake: Assume the AI knows who you are, who the audience is, and what you know.
Why it fails: AI agents lack Theory of Mind. They cannot infer that a "summary" for a CEO is different from a "summary" for a Developer.
Anti-Prompt: "Explain how a car engine works."
Correction (Inversion): Prime the Persona and Audience. "Act as a senior mechanical engineer. Explain how a car engine works to a 5-year-old using analogies involving toys."
Related Strategy: Theory of Mind Prompting — Explicitly define the persona (sender) and the audience (receiver) to adjust complexity.
The "Chaos Agent" (Structure & Format)
The Mistake: Let the model choose the format. Provide zero examples (zero-shot).
Why it fails: You get whatever format is statistically most common in the training data (usually unstructured prose), making it impossible to parse the output programmatically.
Anti-Prompt: "Write a short story. Put it in a table or something if you want."
Correction (Inversion): Force the schema. "Write a story in exactly 3 sentences. Output the result strictly as a JSON object with keys title and story. Here is an example..."
Related Strategy: Explicit Intent Protocol — Enforce strict, machine-readable schemas (JSON/XML) to prevent entropy.
The "Miscast Actor" (Role Mismatch)
The Mistake: Asking a creative "Architect" agent to do rote data entry, or asking a rigid "Executor" agent to plan a strategy.
Why it fails: Mismatch in ambiguity tolerance. The Executor crashes on vague instructions; the Architect hallucinates complexity ("boredom error") on simple tasks.
Anti-Prompt: To a rigid SQL-Executor Agent: "Look at the data and tell me what's interesting about our users."
Correction (Inversion): Route by Ambiguity Tolerance. To the Executor: "Run query SELECT * FROM users WHERE active=true." To the Architect: "Analyze the user table schema and suggest 3 queries to measure retention."
Related Strategy: Role-Based Routing — Assign tasks based on the agent's functional role (Executor vs. Architect).
The "Kitchen Sink" (Overloading)
The Mistake: Mix multiple distinct tasks (explain, code, translate) in a single massive block of text.
Why it fails: Confuses the attention mechanism ("Lost in the Middle" phenomenon). The model often skips instructions buried in the center.
Anti-Prompt: "Explain quantum physics, write a poem about cats, and give me 10 business ideas. Also, translate the explanation to Spanish."
Correction (Inversion): Decompose the Chain. Break the complex goal into atomic steps: "Explain quantum physics, "Translate that explanation."
Related Strategy: Chain of Thought Decomposition.
The "Visual Vibe Check" (The Aesthetic Trap)
The Mistake: Asking the model to "Check if this code is good" or relying on the visual cleanliness of the output (formatting, variable names) as a proxy for correctness.
Why it fails: This falls victim to the Principle of Syntactic-Semantic Decoupling. The model will optimize for "vibes"—producing code that looks professional, passes linters, and has great comments—while hiding deep logical flaws or security vulnerabilities that don't "look" wrong.
Anti-Prompt: "Review this code and tell me if it's clean." OR (implicitly) merging code just because it looks like the rest of the file.
Correction (Inversion): Demand Semantic Verification. Force the model to prove functionality, not just style. Ask it to generate a failing test case for the logic before generating the fix, or ask it to "Explain the edge cases where this logic fails."
Related Strategy: The Semantic Validator — Invert the workflow to verify logic via tests before verifying style via review.