← All guides

“Fix it” — fix what?

Three words, two unanswerable questions
Make it better.

Fix what, and better by what measure. This is the shortest possible broken prompt, and it is genuinely common — because it is what you type when the thing is right in front of you.

Instant checks

9 key structural practices evaluated locally

22% Coverage2/9
  • States a concrete task

    Consider

    No clear task verb, or the prompt is very short.

    Recommendation — Open with the action you want — summarize, extract, rewrite, classify — so the model is not guessing at the job.
  • Avoids vague language

    Consider

    Contains vague words such as "good", "better", or "appropriate".

    Recommendation — Replace judgement words with observable criteria. "Make it better" gives the model nothing; "cut it to 100 words and drop the passive voice" does.
  • Specifies an output format

    Consider

    The wanted output format, length, or structure is never stated.

    Recommendation — Say what shape the answer should take — JSON, a table, three bullets, under 200 words. Left unsaid, models default to prose you then have to parse.
  • Gives context or audience

    Consider

    No audience, purpose, or background is given.

    Recommendation — Say who the output is for and what it will be used for. The same request answered for a beginner and for a specialist are different answers.
  • Avoids dangling references

    Consider

    Uses "it", "this", or "that" with nothing concrete to point at.

    Recommendation — Name the thing. The model cannot see what you are looking at, so a pronoun with no antecedent is a coin flip.
  • Includes an example

    Consider

    No example of the output you want.

    Recommendation — One worked example pins down format, tone, and edge-case handling far more reliably than describing them in prose.
  • Separates data from instructions

    Passed
  • Avoids over-aggressive phrasing

    Passed
  • Defines what a good answer looks like

    Consider

    No statement of what a successful answer contains or avoids.

    Recommendation — Name the constraints — what to include, what to leave out, what would make the answer wrong. It gives the model something to check itself against.
A dangling pronoun, a vague word, and no task at all.

The pronoun problem

You have a document open, a file selected, a paragraph highlighted. "It" is obvious. It is the thing you are staring at.

The model is not staring at anything. It has your text and nothing else. A pronoun with no antecedent in the prompt is not shorthand — it is a blank the model has to fill in, and it will fill it in with whatever was most recently mentioned, which may not be what you meant.

In a long conversation this gets worse rather than better. By turn fifteen there are half a dozen candidate referents for "it", and the model's guess is now a coin flip weighted by recency. You will get a confident answer about the wrong thing, which is harder to spot than an obvious error.

Where it bites hardest

Multi-turn chats. "Now make it shorter." Shorter than what — the last draft, or the original? Both are in the conversation.

Tool-using agents. "Update it with the new value." The agent has read four files. It will pick one.

Reused prompt templates. A prompt that worked when you wrote it — with the document freshly in mind — becomes ambiguous three weeks later when it runs against different input.

The fix is boring and it always works

Name the thing.

Named object, stated bar
Write a launch post announcing Acme Sync.

### Audience
Frontend developers at small startups who have lost work to version mismatches.

### Context
Acme Sync keeps Figma files and production code in step. Nobody in the audience has heard of it.

### Requirements
- Length: About 600 words.
- Output format: Markdown, with a heading per section.
- Avoid: marketing superlatives.

### Example of the wanted output
## Title
Opening line naming the specific pain.

Instant checks

9 key structural practices evaluated locally

100% Coverage9/9
  • States a concrete task

    Passed
  • Avoids vague language

    Passed
  • Specifies an output format

    Passed
  • Gives context or audience

    Passed
  • Avoids dangling references

    Passed
  • Includes an example

    Passed
  • Separates data from instructions

    Passed
  • Avoids over-aggressive phrasing

    Passed
  • Defines what a good answer looks like

    Passed
Every applicable check passes.

Nothing clever happened here. The pronoun was replaced with a noun, and "better" was replaced with the things that would actually make it better. That is the entire technique.

Not every pronoun is a problem

Worth being precise, because an over-eager rule here would be worse than no rule.

"Summarize the article and then critique it" is fine. "It" has an antecedent four words earlier, and no human or model would misread it.

"Make it clear that the deadline is Friday" is fine too — that "it" is grammatical filler, not a reference to anything.

The problem is specifically a verb followed by a bare pronoun with no antecedent anywhere in the prompt: fix it, improve this, rewrite that. That is what our checker looks for, and it took a bug fix to get there — an earlier version flagged any sentence ending in "it.", which meant it fired on the perfectly good "…and then critique it." A checker that cries wolf on good prompts is worse than no checker, so we narrowed it.

The test

Read your prompt as though you have never seen your screen. Every "it", "this", "that", and "the file" — can you point at what it means using only the words in the prompt?

If not, you have not written a prompt. You have written a note to yourself.

Every check in this article ran live in your browser. The same nine run on any prompt you paste in, for free.