A mediocre prompt and a strong prompt often differ by four lines, not by "prompt engineering magic." The skill is diagnosing which line failed, not throwing the whole request into a rewriter and hoping.
People who get stuck in rewrite loops usually change wording without changing structure. They go from "write a good email" to "write a professional, compelling, high-quality email." The model still does not know who it is, what facts to use, what "done" looks like, or what to avoid.
Diagnose before you rewrite
| Symptom in the output | Likely broken part of the prompt | Fix to try first |
|---|---|---|
| Generic tone, could be any brand | Missing role or audience | Add who the model is and who the reader is |
| Made-up product features or stats | Missing inputs / facts | Paste real constraints, numbers, names |
| Right idea, wrong shape (essay when you needed bullets) | Missing format | Specify length, structure, and delivery medium |
| Almost right but hits a known pet peeve | Missing negative constraints | Name the failure mode explicitly |
| Good once, then drifts on retry | No evaluation checklist | Add 3 must-pass checks for the final answer |
If you cannot point to a row, you are not optimizing - you are roulette-spinning synonyms.
Worked example: a sales follow-up that kept coming out soft
Original prompt:
"Write a follow-up email to a prospect who didn't reply after the demo."
Typical output: 180 words of "just checking in," "circling back," and a vague offer to "answer any questions."
Line-by-line diagnosis:
1. Role: absent → model defaults to generic helpful assistant.
2. Inputs: no product, no demo detail, no prospect context.
3. Format: no length or structure.
4. Constraints: no ban on "just checking in."
5. Success check: none.
Optimized prompt:
"You are an AE at a warehouse inventory SaaS. Write a 90-word follow-up to Jordan, ops lead at a 4-location retail brand. Demo was Tuesday: they liked cycle-count automation, worried about onboarding time for store managers. Offer a 20-minute walkthrough of the 3-day store rollout plan only. Do not use 'just checking in,' 'circling back,' or 'bumping this.' One CTA: reply with a time this week or next. Subject line under 6 words."
Same model. Different structure. The second email references the demo concern and makes one ask.
The five-line skeleton that fixes most mediocre prompts
Copy this skeleton and fill every blank. If a blank is empty, that is your diagnosis.
1. Role: You are [specific role] writing for [specific reader].
2. Job: Produce [exact deliverable], not "help with [topic]."
3. Inputs: Use only these facts: [paste]. Do not invent metrics.
4. Format: [length], [structure], [tone].
5. Checks: Before finishing, verify [3 binary checks].
Example checks for a blog outline: "includes one table," "no section titled Introduction," "every H2 names a concrete action."
Optimization is a loop, not a monologue
1. Run the weak prompt once. Save the output.
2. Mark the first sentence that fails. Map it to the table above.
3. Change only the matching line of the prompt.
4. Re-run. Compare.
5. Stop when the failure mode is gone - do not keep "improving" tone forever.
Teams that batch five stylistic tweaks at once never learn which change actually mattered.
Tools that force the diagnosis step
The [Prompt Optimizer](https://auth.mane.dev?site=aiagentstudio&template=prompt-optimizer) on aiagentstudio.pro asks for the current prompt, what went wrong, and the goal - then returns a diagnosis plus a fixed version, not just a prettier rewrite.
If you are still learning the patterns (role, inputs, format, constraints), the [Prompt Engineering Crash Course](https://auth.mane.dev?site=aiagentstudio&template=prompt-engineering) walks through the five patterns with before/after examples on your own use cases.
Common false fixes
- Adding "be creative" or "think step by step" when the real problem was missing inputs.
- Stacking 12 constraints until the model cannot satisfy them all and starts ignoring half.
- Optimizing for eloquence when the user needed a table they could paste into Notion.
- Trusting the optimized prompt without a second test on a new input set.
Over-constraining is as common as under-specifying. If the model keeps failing, remove two low-priority constraints before adding three more.
A 10-minute practice drill
Take one prompt you used this week that produced something "fine but not usable." Paste it into a doc. Label each sentence Role / Input / Format / Constraint / Noise. Delete Noise. Fill empty labels. Re-run once.
That single pass is prompt optimization. Everything else is branding.
When the diagnosis is clear, the rewrite is short. When the diagnosis is fuzzy, no amount of "make it better" language will save the output.
Mini case: content brief that kept producing fluff
A marketing manager had this prompt:
"Create a content brief for a blog about remote work tools."
Output: a generic outline with "Introduction," "Benefits," "Challenges," "Conclusion," and no audience, no keyword, no deliverable definition.
Diagnosis mapped to the table: missing role, missing inputs, missing format.
Fixed version:
"You are a content lead at a 25-person B2B SaaS company selling async standup software. Audience: engineering managers at 50–200 person product companies. Keyword intent: informational, 'async standup best practices.' Produce a one-page brief with: working title, search intent in one sentence, 5 H2s that are action-named (not Introduction), 3 competitor angles we should not copy, required proof points I must supply (list as blanks), and a success definition: reader can run a 2-week async standup pilot after reading. 250 words max for the brief itself."
After one retest, the brief had blanks for proof points instead of invented stats. That is optimization: change the structure that caused the failure mode.
When not to optimize further
If the model already returns the right structure and you are only polishing adjectives, stop. Hand the draft to a human. Endless micro-edits to the prompt for tone are a trap that burns more time than writing the final paragraph yourself.