You asked “what is the best way to scale a startup engineering team?” and got 800 words covering hiring, culture, process, tooling, and remote work. The answer is technically right and applies to literally every startup that has ever existed. It does not apply to yours.
Fastest fix: add scope (a number), constraints, the decision-maker, and a one-line success criterion, then swap the open verb (“how should I”) for a decision verb (“pick one of these three and defend it in three sentences”). That single rewrite is usually enough. The rest of this page is the full checklist for when it is not.
Broad questions produce broad answers because the model has to write something defensible across the entire space of inputs your question allows. When that space is huge, the only safe output is the average of common advice, which is mediocrity by definition. This is not a model limitation you can “prompt harder” past with GPT-5.5, Claude Opus 4.7, or Gemini 3.1 Pro — a smarter model just writes a more polished average. The fix is to shrink the input space until one answer is correct. Anthropic frames the same idea as context engineering: useful context is not more text, it is the small set of details that changes what a good answer looks like (audience, goal, constraints, priorities).
Which bucket are you in?
Run your prompt through this table before rewriting. Most broad prompts fail on two or three rows at once.
| Missing element | Tell-tale sign in your prompt | One-line fix |
|---|---|---|
| Scope / scale | No numbers (team of 3 or 300?) | State the size, volume, or stage |
| Constraints | No “given X” or “subject to Y” | Add budget, deadline, stack, headcount |
| Decision-maker | No role named | Name who decides and what they optimize for |
| Success criterion | No “good output looks like…” | Define success in 2-3 measurable terms |
| Input data | Zero concrete numbers, names, or files | Paste 5+ specific facts about your case |
| Decision verb | Asks “how” or “what’s best,” not “pick / rank / choose” | Force a choice from named candidates |
Common causes (expanded)
1. No scale or scope
“How should I structure my engineering team” can mean a team of 3 or 300. The model writes the answer that covers both, which fits neither.
How to spot it: your question has no numbers.
2. No constraints
“What’s the best database” with no traffic, no cost limit, no team familiarity. The model has nothing to eliminate options against, so it lists them all.
How to spot it: the question is open with no “given X” or “subject to Y” clause.
3. No deciding stakeholder
“What’s the right choice for our roadmap” — for whom? The CEO has different criteria than the engineering lead. Without a stakeholder, the model averages across all of them.
How to spot it: no role named in the prompt.
4. No success criterion
“Help me think through X” with no statement of what counts as helpful. The model writes a survey because surveys are universally non-controversial.
How to spot it: prompt has no “good output looks like” clause.
5. No input data
You asked for advice with no specifics about your situation, so the model has to write generically.
How to spot it: zero concrete numbers, names, files, or data points in the prompt.
Shortest path to fix
Step 1: Replace open verbs with decision verbs
Bad: "What's the best way to scale our engineering team?"
Good: "We're 8 engineers shipping a B2B SaaS, currently 2 squads,
lead time from PR to prod is 4 days. We want to halve lead time
within 1 quarter. Pick ONE intervention from this list, defend in 3 sentences:
(a) move to trunk-based development
(b) add a dedicated platform engineer
(c) cap PR size at 200 lines"The decision verb (“pick”), the named candidates, and the criterion together force a concrete answer. “Pick / rank / choose from these options” is far harder to hedge than “how should I” or “what’s best.”
Step 2: Add 5 lines of concrete context
Stack: <runtime, framework, key deps with versions>
Scale: <users, requests, data size>
Constraints: <budget, deadline, team>
Tried: <what already failed>
Goal: <the deliverable with a success criterion>This template converts a generic prompt into a specific one. Front-load these constraints so the model never has to invent assumptions — practitioner write-ups in 2026 consistently report that naming the audience and use case roughly halves the number of follow-up iterations needed.
Both vendors recommend labeling your sections rather than writing one paragraph. Models process a structured prompt more reliably than a wall of text, so use plain headers like CONTEXT:, TASK:, and FORMAT:, or XML-style tags (<context>…</context>) that Anthropic recommends for Claude.
Step 3: Name the decision-maker
"This decision will be made by the eng lead and reviewed by the CTO.
The eng lead optimizes for delivery speed. The CTO optimizes for retention risk."A stakeholder plus their priority calibrates the answer toward one reader instead of an average of all readers.
Step 4: Define success in numbers
Success criteria for this answer:
- One concrete intervention named (not a list).
- Defense in under 100 words.
- One risk identified and mitigated.
- One leading indicator we can watch in 2 weeks.Measurable criteria stop the model from hedging back into prose. This is the single most reliable lever, and it is the one most prompts omit.
Step 5: If still vague, ask the model what it needs
"What 3 specific data points would you need from me to give a
concrete answer instead of a generic one?"Answer those, paste them back, and re-ask. A two-turn approach beats a single-shot vague prompt almost every time, because the model itself surfaces the missing context.
Step 6: Split broad into narrow (prompt chaining)
If the question really is huge (“how do we scale”), split it. Decomposing one broad prompt into a chain of narrow ones — where each output feeds the next — is a documented technique that measurably outperforms single-shot prompting on multi-step problems, because each step has a clear objective and you can validate it before continuing:
- Prompt 1: identify the top 3 bottlenecks given our specific data.
- Prompt 2: for the top bottleneck, pick the intervention.
- Prompt 3: design the rollout for that intervention.
Three narrow prompts produce far more actionable output than one broad prompt, and you catch a wrong turn at step 1 instead of in a 800-word wall.
How to confirm the fix
- The answer names a specific intervention, not a list.
- The answer references your concrete facts (it would not make sense for a different company).
- A different team with different facts would get a different answer — if not, your prompt is still too broad.
- A teammate can act on the answer without follow-up questions.
- Output word count is concentrated on the recommendation, not on a survey of options.
If it still fails
- Context may be missing — add more specifics about your situation.
- Tell the model what to ignore (e.g., “ignore generic startup advice; assume I already know the basics”).
- Force a binary choice (“A or B”); a binary is harder to hedge than an open question.
- If the question truly has no answer with your current data, your bottleneck is data collection, not prompting — no rewrite fixes that.
FAQ
Why does a smarter model not fix a broad prompt? Because the problem is the size of the input space, not the model’s intelligence. GPT-5.5 or Claude Opus 4.7 will write a more eloquent average, but it is still an average. Narrowing the question is the only lever that changes the kind of answer you get.
How specific is specific enough? Specific enough that swapping in a different company’s facts would change the recommendation. If the same answer would satisfy any reader, you have not narrowed it. Aim for at least one number (scale), one constraint, and one named decision-maker.
Is a long prompt the same as a specific prompt? No. Length and specificity are different axes. A long prompt can still be vague, and a padded prompt can bury the constraints that matter. Add details that change the answer, not volume. If your prompt is already long and still vague, see Long prompt degrades output.
Should I narrow up front or let the model ask me? Front-load the constraints when you know them — it is faster and avoids wasted turns. When you are not sure what is missing, use Step 5 and let the model tell you which three data points it needs.
What if I genuinely want to explore an open topic? Then ask for one concrete strawman first (“propose one opinionated plan, even if imperfect”), and iterate against it. A strawman gives the model a single position to defend, which is more useful than a balanced survey.
Prevention
- Default to narrow prompts. Always include scope, constraints, decision-maker, and success criteria.
- Use a pre-send checklist: who decides, with what data, optimizing for what, and what good output looks like.
- For exploration, ask for one concrete strawman first, then iterate against it.
- Audit your broad-question habits: every “what’s the best way to” should trigger a narrowing step.
- For team workflows, build a “narrow prompt” template covering scope, constraints, and a decision verb.
- When tempted to ask broad, ask “what’s the smallest concrete decision I can ask about?” instead.
For deeper background, Anthropic’s effective context engineering write-up and OpenAI’s prompt engineering guide both make the same point: specific, structured context beats clever wording.
Related reading
- AI answer too vague
- Unclear task boundary
- Prompt asks “best” undefined
- No success criteria
- Prompt missing audience
Tags: #Troubleshooting #Prompt #Prompt quality #Vague answer