Articles tagged with #Prompt engineering
Listed 'text, watermark, extra fingers' in the negative prompt and got all three anyway? It is usually one of seven causes: not wired up, CFG too low, contradicted by the positive prompt, truncated past 75 tokens, wrong weight syntax, an overtrained checkpoint, or a Flux model that ignores negatives by design.
You asked for a slow dolly-in and got a dolly-out, or pan left came back as pan right. Fix it with start/end framing, screen-space language, and the built-in camera sliders in Runway, Kling, and Flow.
You gave the model 5 examples; 2 are great, 3 are mediocre, and it averages toward the mediocre ones. Why example variance hurts and how to curate down to 3-5 consistent ones.
You asked for JSON matching a schema. Most calls return valid JSON, some return prose with JSON inside, some omit fields. Description vs enforcement, and how to fix at the API layer.
The model produced a citation like Smith et al. 2019 and the paper does not exist, or a URL that 404s. Why citation hallucination happens and how to stop it for good.
You prompted in English and the model answered in Chinese, or it switched mid-output. The exact cause of language drift and the system-prompt + retry pattern that locks the output language, verified June 2026.
The model's reply ends mid-sentence, mid-JSON, or with an unclosed code block. It is almost always the token cap. How to size it, detect truncation per SDK, and recover.
You asked for 10 ideas and got 3, or 10 slots padded with filler. Why list-length prompts under-deliver, and the prompt + schema fixes that actually get you N distinct items.
Your prompt still says 2023 in 2026, so the model recommends GPT-4, quotes old pricing, and cites dead frameworks. Fastest fix: inject the current date dynamically. Plus a diagnosis table and how to confirm it's gone.
You asked the model to write, refactor, or draft something and got a 10-bullet plan instead. Here is why it switches to advice mode and the exact prompt edits that force a finished artifact.
You listed five rules. The model honored four and quietly dropped the one that mattered most. Here is why constraints get dropped and how to make the critical one stick.
Your criteria are vague — "engaging", "professional", "innovative" — so the model interprets freely. Turn each adjective into a testable rule with a 10-second check.
When two prompt rules fight, the model averages them into something nobody asked for. Rank your constraints so it knows which to drop.
A casual aside at the end of your prompt can overwrite the careful rules at the top. Anchor the hard rules at both ends so the last line stops winning.
Describing a tone or shape in words makes the model approximate; pasting one concrete example makes it match. How to pick, place, and structure 1-5 examples to lock the output you want.
"Do not be generic" tells the model what not to do without telling it what to do, so it dodges the word and keeps the behavior. Pair every 'do not' with a concrete 'do'.
You wanted a structured answer and got 600 words of paragraphs. Here is how to force clean JSON or a fixed template, every run.
When a prompt has no success criteria, "good" defaults to whatever the model thinks sounds confident. Here is the 5-line success block that ends revision purgatory.
Run the same "what's best?" prompt three times, get three answers. Replace "best" with an axis, weights, and a tie-breaker to get one defensible pick.
A prompt that worked elsewhere produces nonsense for your current task because the old audience, format, and examples are still baked in. Here's how to rebuild from the goal.
"Please make it amazing!" feels persuasive, but it tells the model nothing to act on. Swap adjectives for checkable rules.
You pasted everything as a flat block, so the model can't tell critical lines from background. Add labels, tag your sources, and put the task where attention is highest.
You attached three documents and the model treated them all as equally authoritative — including the rejected draft. Here's how to label provenance so the right source wins.
No audience in the prompt means the model writes for an imaginary average reader and pleases no one. Fix it with a one-line audience block that calibrates vocabulary, depth, and tone.
"You are a senior engineer" sets the tone but does not change the answer. Research says expert personas rarely raise accuracy; rules, format, and examples do.
Stacking 5+ examples makes the model copy whichever one resembles your input instead of executing the task. Cut to 1-3, and on reasoning models try zero-shot first.
Stacked five tasks in one prompt and got one good answer, one weak one, and three half-finished? Here is how to split the work so every task lands.
You asked the model to fix one function; it also reformatted two others and renamed a constant. Draw an explicit in-scope / out-of-scope boundary so the edit holds.