Few-Shot Examples Have Uneven Quality and Drag Output Down
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.
🧠 AI Prompt Quality Issues
Prompt-side problems that degrade output: clarity, structure, output control, and behavior fixes — with Before / After examples.
About 70% of "bad AI answers" are prompt-structure problems, not model problems. This hub is organized into four tracks: (1) Clarity — long prompts get worse, conflicting instructions, too-broad prompts, unclear task boundaries, too many tasks in one prompt, background that buries the task, emotional wording instead of operational instructions; (2) Structure — no output format, missing or too many examples, role instruction alone, missing context hierarchy, missing source hierarchy, prompts copied from another task; (3) Output Control — vague answers, polished-but-not-actionable output, lists instead of execution, no success criteria, ambiguous evaluation criteria, dropped constraints, recency overrides earlier rules, vague negatives, mixed-tone instructions, style vs format conflicts; (4) Behavior Repair — hallucinations, gap-filling assumptions, style drift, rewritten key logic, over-editing on light rewrite, missing decision rules, missing audience, unexpected refusals. Every article ships a "bad prompt → good prompt" comparison and at least 5 concrete repair techniques you can apply to your next prompt immediately.
Common problems
- Long prompt makes the answer worse Compress the goal to one sentence; section the body; end with format.
- Conflicting instructions weaken AI output Pick the winner; put it in a non-negotiable block; demote others.
- Prompt too broad to produce useful results Add scope, constraints, audience, decision rule.
- AI changes scope because the task boundary is unclear Closed verbs + explicit out-of-scope list + demand a diff.
- Prompt asks for too many tasks at once 1 prompt = 1 task. Three jobs = three turns.
- Long background hides the actual task Imperative deliverable first; move background under headings.
- Prompt uses emotional wording instead of operational instructions Turn adjectives into measurable rules + positive/negative examples.
- No output format specified End with a schema block; forbid extra prose.
- Missing examples leads to output drift One positive + one negative example beats five paragraphs.
- Too many examples overwhelm the task 1-3 representative examples; task instruction after them.
- Role instruction alone is not enough One-line role; rules / schema / examples carry the load.
- Prompt lacks context hierarchy Section into Task / Background / Constraints / Reference / Output.
- Prompt lacks source material hierarchy Label each source with date+role; declare canonical explicitly.
- Prompt copied from another task fails here New task = rebuild; do not retrofit an old prompt.
- AI answer is too vague Add 5 lines of context; ask for a decision; ban hedging words.
- Output sounds polished but is not actionable Force artifacts: paths, commands, numbers, owners.
- AI gives lists when execution is needed Imperative + "Do not list steps. Produce the result."
- Prompt does not state what success looks like End with 3-5 measurable acceptance checks.
- Ambiguous evaluation criteria cause weak answers Turn adjectives into testable rules + pass/fail examples.
- AI ignores the most important constraint Hoist the critical rule to the top + self-check at the end.
- Model follows the latest sentence and ignores earlier rule Restate task + hard rules at the end; no "by the way" notes.
- Negative constraints are too vague Pair each "do not" with a "do"; keep banned list short.
- Mixed tone instructions create unstable output One primary tone; other tones as nuances only.
- Requested style conflicts with requested output format Schema for systems, markdown for humans, split into two passes if needed.
- AI hallucinated facts Provide the source + "Write UNCERTAIN if unsure."
- Model fills missing details with unwanted assumptions "Write UNKNOWN and stop"; never invent specifics.
- AI output style keeps drifting Replace adjectives with hard rules; anchor with examples; pin model+temp.
- AI rewrote a critical section you did not want touched Do-not-modify list + demand diff, not a full rewrite.
- AI over-edits when only a light rewrite was requested "edit" not "polish"; cap change %; demand diff.
- Prompt asks for "best" without defining a decision rule "Best on <axis>" + weights + tie-breaker.
- Prompt does not define the intended audience One line: role, expertise level, current task.
- Prompt triggered an unexpected refusal Reframe with legitimate context; swap trigger words for neutral ones.
- Prompt mixes system and user instructions Global rules → system; current task → user.
In-depth articles in this hub
Model Returns Invalid JSON Because Schema Was Described, Not Enforced
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.
Model Invented Fake Citations and URLs
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.
Model Replies in the Wrong Language (How to Lock It)
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.
LLM Response Cut Off Mid-Sentence: max_tokens Too Low (2026 Fix)
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.
Prompt Asks for 10 Items, Model Returns 3 and Stops
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.
Prompt Has a Stale Year That Anchors the Model to Wrong Context
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.