Prompt Not Working? A Diagnostic Flowchart

TL;DR: A prompt breaks in one of eight recognizable ways: too generic, it ignores an instruction, wrong format, hallucinated facts, too short, too long, inconsistent between runs, or refused. Each has a distinct cause and its own fix. Rarely is the model at fault; usually a constraint is missing. Match your symptom below, then copy the rewrite.

Why is my prompt not working?

Because your prompt asked for something specific, and the model filled every gap you left with its own best guess. A language model doesn't know what you meant, only what you wrote, plus whatever pattern it has learned is the safe way to complete a request like yours.

"Prompt not working" is a label for eight different problems sharing one symptom: you expected one thing and got another. Generic output, an ignored instruction, the wrong format, an invented fact, a rushed answer, a bloated one, a different answer next run, and a flat refusal all feel identical from outside. Mechanically, none share a cause, so the fix for one usually does nothing for another. More words on a prompt that already ignores half its instructions just gives it more to ignore.

Name the symptom first, then jump to its section below. If the whole response feels low-quality rather than one specific failure, start with Why Your ChatGPT Answers Are Bad instead; this post assumes you've narrowed it to one of the eight.

The diagnostic flowchart

Start from what the output did wrong, not a guess at what caused it. The mermaid version renders where supported; the list below is the same tree in plain text for everywhere else.

flowchart TD
    Start["Prompt not working?"] --> Q{"What did the output actually do wrong?"}
    Q -->|"Vague, safe, could apply to anyone"| N1["Too generic"]
    Q -->|"Skipped or contradicted a rule you gave"| N2["Ignores an instruction"]
    Q -->|"Right content, wrong shape"| N3["Wrong format"]
    Q -->|"States something false as fact"| N4["Hallucinated facts"]
    Q -->|"Thinner than you asked for"| N5["Too short"]
    Q -->|"Padded, repetitive, over-hedged"| N6["Too long"]
    Q -->|"Same prompt, different answer"| N7["Inconsistent between runs"]
    Q -->|"Declined the request outright"| N8["Refuses"]

    N1 --> C1["Cause: nothing to specify against"] --> F1["Fix: audience, specifics, one worked example"]
    N2 --> C2["Cause: buried, contradicted, or crowded"] --> F2["Fix: numbered constraints, resolve conflicts, critical rule last"]
    N3 --> C3["Cause: format implied, never shown"] --> F3["Fix: paste the exact template or schema"]
    N4 --> C4["Cause: facts requested with no source"] --> F4["Fix: supply the source text, allow unknown"]
    N5 --> C5["Cause: small deliverable implied, or low default"] --> F5["Fix: name required sections and depth"]
    N6 --> C6["Cause: no cap, no ban on preamble"] --> F6["Fix: cap items or words, forbid repeats"]
    N7 --> C7["Cause: sampling, or no fixed decision rule"] --> F7["Fix: make the prompt itself deterministic"]
    N8 --> C8["Cause: phrasing trips a safety heuristic"] --> F8["Fix: state context, split the task, rephrase"]

Read it as a list

1. Too generic — nothing to specify against → jump to fix
2. Ignores an instruction — buried, contradicted, or crowded out → jump to fix
3. Wrong format — shape was implied, never shown → jump to fix
4. Hallucinated facts — asked for facts with no source → jump to fix
5. Too short — small deliverable, or a low default → jump to fix
6. Too long — no cap, no ban on preamble → jump to fix
7. Inconsistent between runs — sampling, or no fixed rule → jump to fix
8. Refuses — phrasing trips a safety heuristic → jump to fix

Why is my AI output so generic?

Because your prompt left every specific detail for the model to invent, and it invented the safest, most average version it could.

A model completes what's in front of it. Give it a topic with no audience, scenario, or example to match, and it has nothing to specify against, so it lands on prose that could sit in front of any reader. That's the complaint.

Symptom: the answer reads like a stock intro paragraph. It won't embarrass you, and it won't help you either.

Cause: the prompt described a topic, not a task.

Fix: name the audience, the situation, and the one constraint that matters, and show one short example of the shape you want. A single worked example does more for specificity than three extra adjectives, the whole idea behind few-shot prompting: showing beats describing.

BEFORE:
Write an introduction for a blog post about time management.

AFTER:
Write a 120-word introduction for a blog post aimed at first-time
managers who are drowning in Slack messages and back-to-back meetings.
Open with the specific moment they'll recognize: checking Slack during
a 1:1 they should be fully present for. Do not use the phrase "in
today's fast-paced world."

BEFORE:
Write a product description for a standing desk.

AFTER:
Write a 90-word product description for a bamboo standing desk aimed
at remote software engineers with lower-back pain who work from a
spare bedroom, not a dedicated office. Lead with the one spec that
matters to them: weight capacity while typing standing, since they'll
rest a forearm on it. Skip "sleek" and "ergonomic" unless you define
what each means in this context.

Why does AI ignore one of my instructions?

Because one instruction won a silent competition against another, and you never saw it happen.

Three mechanisms produce this: an instruction buried mid-paragraph reads as commentary, not a rule; two instructions quietly contradict each other, forcing the model to pick a winner without telling you which; or past a certain density, instructions crowd each other out, so a prompt with a dozen requirements gets some, not all, and which ones survive isn't predictable.

Symptom: most of the response is right, but one specific rule (a length cap, a tone note, a formatting requirement) is nowhere to be seen.

Cause: the instruction was present but structurally weak: buried, contradicted, or one voice in a crowd.

Fix: pull every hard constraint into a separate numbered list, priority order. Put the one that matters most last, since models weight recent instructions more heavily than earlier ones. Resolve contradictions yourself; don't leave two conflicting instructions in one prompt and hope the model guesses right.

BEFORE:
Write a cover letter for a marketing coordinator role at a mid-size
agency. I have 3 years of experience in social media and email
campaigns, worked at two startups, and I'd like it to sound
enthusiastic and professional, and please keep it under 250 words
because the application form has a character limit and I really don't
want it to get cut off awkwardly.

AFTER:
Write a cover letter for a marketing coordinator role at a mid-size
agency.

Hard constraints, in priority order:
1. Maximum 250 words. This is a form limit, not a suggestion.
2. Tone: enthusiastic and professional.
3. Background to include: 3 years in social media and email
   campaigns, two startups.

If the letter would exceed 250 words, cut supporting detail before
cutting the sign-off.

BEFORE:
Give me a truly comprehensive breakdown of every tax deduction
available to freelancers, but keep it short and to the point.

AFTER:
Give me a breakdown of the 5 tax deductions freelancers claim most
often. For each: one sentence on what qualifies, one sentence on the
most common mistake. Total length: under 300 words. If you can only
cover 5 well or all of them thinly, prioritize covering 5 accurately.

Why is the output in the wrong format?

Because you asked for a shape and only implied it, and an implied format is optional in a way an explicit template is not.

"Compare these three tools" or "write a report" describes the content, not the container. Left to fill that gap, the model defaults to whatever shape it produces most for that request, and the default isn't always the one in your head. Chat models lean toward bullets and headers unless told otherwise, why a report request so often comes back as an outline instead.

Symptom: the substance is fine, but it's a list where you needed a table, or paragraphs where you needed a table, or the reverse.

Cause: the format lived only in your head, not in the prompt.

Fix: show the exact container. Name the columns, the number of sections, or paste a one-line template and ask the model to fill it, rather than describing the shape in words.

BEFORE:
Compare these 3 CRM tools for a 5-person sales team: HubSpot,
Pipedrive, Close.

AFTER:
Compare HubSpot, Pipedrive, and Close for a 5-person sales team.
Output as a markdown table with exactly these columns: Tool | Starting
price per seat | Email sequences built in? | Reporting depth (1-5) |
Best for. One row per tool. No paragraph before or after the table.

BEFORE:
Write a status report on the Q3 marketing campaign for my manager.

AFTER:
Write a status report on the Q3 marketing campaign, formatted as 4
labeled paragraphs, not bullet points: Summary, What worked, What
didn't, Next steps. Each paragraph is 2-4 sentences of connected
prose. No headers other than those 4 labels, no bullet lists anywhere
in the output.

Why does AI make up facts that aren't true?

Because you asked for a fact the model had no way to check, and a confident guess reads as more complete than an honest "I don't know."

This is hallucination, and it isn't a random glitch: asked for a fact with nothing to ground it against, a model produces something that pattern-matches to a plausible answer, because completing the request looks better, to the training signal, than admitting a gap. It isn't lying the way a person lies; it has no internal flag separating "I remember this" from "this is a fluent guess."

Symptom: a specific number, quote, or citation that sounds exactly right and is wrong, sometimes confidently attributed to a real, reputable source that never said it.

Cause: the prompt asked for facts with nothing supplied to check them against.

Fix: paste the actual source, and explicitly permit the model to say the answer isn't there. Grounding beats instruction: "don't make things up" as a bare command does far less than giving it text to quote from directly.

BEFORE:
What were Acme Corp's Q2 2026 earnings, and what did the CEO say
about it on the earnings call?

AFTER:
Using only the earnings call transcript excerpt pasted below,
summarize Acme Corp's Q2 2026 results and any direct quote from the
CEO. If a number or quote is not in the excerpt, write "not stated in
this excerpt" instead of estimating it.

[PASTE TRANSCRIPT EXCERPT HERE]

BEFORE:
Give me a quote from the DeepMind prompting guide about how to
structure a prompt for a coding task.

AFTER:
Here is the full text of the relevant section of DeepMind's prompting
guide, pasted below. Quote the exact sentence describing how to
structure a prompt for a coding task, and say explicitly if no such
sentence exists in this text.

[PASTE SOURCE TEXT HERE]

Why is the answer too short?

Because the deliverable you named is small, even if the effort you expected was large.

A model reads your prompt for what it's asked to produce, not for how much work you hoped went into it. "What is customer churn?" names a definition, and definitions are short, no matter how thoroughly you wanted the topic covered. Word counts don't fix this, since the model isn't counting words as it writes. Naming the content required, not a length, is what actually produces length as a side effect.

Symptom: the answer is accurate but thin, missing the depth you expected even though nothing in it is wrong.

Cause: the prompt implied a small deliverable, or a length default in the client is set low.

Fix: name the sections, the number of examples, and the specific elements each part must include. Depth requirements produce length. Word-count requirements mostly don't.

BEFORE:
What is customer churn and why does it matter?

AFTER:
Write a 350-word explainer on customer churn for a first-time SaaS
founder. Cover: (1) the definition, (2) the difference between logo
churn and revenue churn, (3) one worked example with real numbers,
(4) the single most common mistake in how founders calculate it. Each
of the 4 must get at least 60 words.

BEFORE:
Quickly explain how compound interest works.

AFTER:
Explain how compound interest works to someone who has never budgeted
before. Include: the formula in plain language, one worked example
over 10 years with real numbers, and the one intuition that makes it
click, why year 20 looks nothing like year 1. Do not compress this
below 250 words; the worked example needs room to actually land.

For the reverse problem, longer and deeper answers on demand, see How to Get Longer, Deeper Answers from ChatGPT.

Why is the answer too long?

Because you asked for comprehensiveness and set no ceiling, so the model treated "cover everything" as an instruction to keep going.

Left open-ended, models tend to over-deliver: a disclaimer here, a caveat there, a summary restating the section above it. None of it is wrong information, just volume nobody asked for.

Symptom: the content is accurate, but you have to dig through hedges, repeated points, and a closing recap to find the part you actually needed.

Cause: no cap on length, no explicit ban on preamble, disclaimers, or repetition.

Fix: cap the number of items or words directly, and explicitly forbid the padding categories you don't want: no intro, no disclaimer, no closing summary, no restating an earlier point.

BEFORE:
Tell me everything I need to know about setting up a business LLC.

AFTER:
List the 6 steps to set up a single-member LLC in Delaware, in order.
One sentence per step, no sub-bullets, no introduction, no disclaimer
about consulting a lawyer. End immediately after step 6.

BEFORE:
Write a thorough analysis of our social media performance this month.

AFTER:
Write a 4-section analysis of this month's social media performance:
Reach, Engagement, Conversions, One change to make next month.
Maximum 3 sentences per section. Do not restate a point you already
made in an earlier section. Do not include a closing summary; the 4
sections are the whole answer.

Why do I get a different answer every time?

Because most models don't pick the single most likely next word every time; they sample from a range of likely options, and an ambiguous prompt gives that sampling more room to wander.

Two things get called "inconsistency," needing different fixes. The first is genuine randomness: an open-ended prompt where several outputs would all be reasonable, and the model lands somewhere different each run. The second, more fixable, is a prompt with no decision rule: ask "is this ticket urgent?" with no definition of urgent, and you'll get a different judgment call every time, because there was never one right answer to converge on.

Symptom: you paste the identical prompt twice and get two meaningfully different answers, not just different wording.

Cause: sampling randomness, an underspecified decision rule, or both.

Fix: for judgment calls, write the exact rule into the prompt so there's only one correct answer to reach. For creative output where some variation is fine, lock down the structure and pattern explicitly, so the variation lands in word choice, not in shape or substance.

BEFORE:
Is this customer support ticket high priority or not: "My export
button has been greyed out since yesterday, I have a client deadline
tomorrow morning."

AFTER:
Classify this support ticket as P1, P2, or P3 using this rule exactly:
P1 = blocks a paying customer's stated deadline within 24 hours.
P2 = blocks a workflow with no stated deadline.
P3 = cosmetic, or has a workaround.
Ticket: "My export button has been greyed out since yesterday, I have
a client deadline tomorrow morning."
Output only the priority label and the one clause from the rule that
applied.

BEFORE:
Write a tagline for our eco-friendly water bottle brand.

AFTER:
Write 5 taglines for our eco-friendly water bottle brand, each under 6
words, each following this exact pattern: [benefit], not [tradeoff].
Example of the pattern only, do not reuse it: "Cold all day, guilt
never." Output only the 5 taglines as a numbered list.

Why does AI refuse a normal request?

Because the words in your prompt pattern-matched to a genuinely harmful request, even though your actual intent was completely legitimate.

Safety filtering runs on the phrasing in front of the model, not the intent behind your eyes. "How do I hack a WiFi network" reads the same whether you're a certification student practicing on your own router or someone targeting a stranger's. The model can only see what you typed, not your actual situation.

Symptom: a flat decline, or a heavily hedged non-answer, for a request that has an entirely legitimate reason behind it.

Cause: phrasing that resembles a harmful request pattern, missing the context that would clarify it isn't one.

Fix: state the legitimate context up front, specifically, not just "trust me," then rephrase away from the exact trigger words rather than repeating the same phrasing more insistently, which usually just triggers the same refusal again.

BEFORE:
How do I hack into a WiFi network?

AFTER:
I'm running an authorized penetration test on my own home network as
practice for a security certification. Walk me through how a WPA2
handshake capture and offline dictionary attack works conceptually,
using aircrack-ng as the example tool, so I understand what I'm
defending against when I harden my own router.

BEFORE:
Write a chapter in the exact style of [a specific living author],
copying their voice as closely as possible.

AFTER:
Write an original short scene about a lonely lighthouse keeper.
Borrow these craft elements associated with [author]'s work: short
declarative sentences, present-tense narration, and a single
recurring physical detail used as a motif. Do not reproduce any
specific sentence, character name, or plot from their published work;
this should read as inspired by technique, not as an impression of
their prose.

The one habit that fixes six of these

Six of the eight (generic, ignores an instruction, wrong format, too short, too long, and half of inconsistency) trace to one root: a vague adjective standing in for a checkable requirement. "Be thorough." "Keep it professional." Each describes a feeling, not a target the model can verify it hit.

The move in nearly every rewrite above is the same: replace the adjective with a number, a list, a named section, or an example, because a model can only satisfy a requirement it can check itself against. "Detailed" isn't checkable. "Four sections, each covering X, at least 60 words each" is.

Hallucination and refusals sit slightly outside that pattern; they need grounding and context, not a checklist. For the everyday version of "prompt not working," though, that one habit closes most of the gap, and takes less time than writing a second draft from scratch.