A weak prompt reads like brainstorming notes. A strong prompt reads like a miniature Product Requirements Document (PRD), where each line earns its existence. If your team wants reliable output, this is non-negotiable.
Start again from prompt design as product design. The core move is simple and hard: stop asking whether a prompt sounds smart, and ask whether each sentence buys a measurable behavior.
Most teams write prompts in one pass. That is backwards. You should write in three passes.
Pass one is intent. What user decision are we trying to improve right now?
Pass two is risk. What can go wrong in this context and how costly is each failure?
Pass three is language. Which sentences create the behavior we want with the least ambiguity?
When teams skip pass one, they get generic prompts. When they skip pass two, they get unsafe prompts. When they skip pass three, they get decorative prompts.
Here is the discipline to apply line by line.
Sentence type one: role framing.
Role framing tells the assistant how to prioritize tradeoffs. “You are a strategist” is weak. “You are an operations advisor helping a first-time manager decide among three options under a two-week deadline” is useful because it encodes pressure and purpose.
Sentence type two: task definition.
Task lines should define one job and one output intent. If a line asks for analysis, rewrite, and coaching simultaneously, you have already lost control. Split multi-intent prompts into explicit stages.
Sentence type three: context payload.
Context is where many teams overstuff. More text is not more clarity. Include only details that change the decision. If a fact does not alter recommendations, it is clutter.
Sentence type four: constraints.
Constraints are where product quality actually lives. Tone, length, forbidden claims, confidence handling, and escalation thresholds belong here. Without constraints, the assistant optimizes for plausible completion, not product trust.
Sentence type five: output contract.
Specify response structure for the user moment. A manager under pressure does not need a philosophical essay. They need a ranked recommendation with rationale and risks. Format is not cosmetic. Format controls usability.
Sentence type six: failure behavior.
This is the line almost everyone forgets. What should the assistant do when context is incomplete or conflicting? If you do not specify this, the model will guess. Guessing in high-stakes contexts is product negligence.
Now the key review question for each sentence: what defect appears if this sentence disappears?
If the answer is “probably nothing,” delete it.
If the answer is “the assistant might overstate confidence,” keep it and tighten wording.
If the answer is “reviewers would interpret this differently,” split it into two lines.
This method changes team dynamics. You stop debating personal preference and start debating failure prevention.
Consider Linear’s AI summary design. The product did not try to sound like a genius analyst. It aimed for quiet utility: useful compression, predictable structure, low interruption cost. That is exactly what sentence-level discipline enables. Every instruction serves a narrow product purpose.
Contrast that with teams who write broad prompts like “summarize this thread with insights and action items and concerns and next steps in a friendly tone.” That single line hides too many decisions. Different reviewers read different requirements. The assistant picks one interpretation and disappoints someone every time.
Another mistake is mixing non-negotiable requirements with preferences.
A non-negotiable requirement is something you can fail in review: “do not invent sources,” “flag uncertainty explicitly,” “keep to five bullets.”
A preference is style seasoning: “sound warm,” “be crisp,” “avoid buzzwords.”
Both matter, but they should not carry equal weight in prompt design. Put non-negotiables first, because they protect user trust.
You also need a hierarchy for conflicts. If two lines clash, which wins?
For example, concision and completeness often conflict. Decide in advance which one has priority for a given user moment. A first-draft ideation tool can prioritize breadth. A decision memo assistant should prioritize completeness with explicit uncertainty markers.
Sentence-level discipline is also how you keep prompts maintainable across model updates. Large language model behavior changes over time. If your prompt is a tangled paragraph, you cannot isolate regressions. If your prompt is structured with clear purpose per line, you can identify exactly what broke.
Use change labels in your workflow:
Intent change: we changed product objective.
Risk change: we discovered new failure mode.
Format change: users found output hard to use.
Tone change: brand or audience shift.
This keeps prompt edits legible over months, not just days.
There is also a hard truth for leaders: vague prompts often signal vague product ownership. If no one owns decision quality at the sentence level, quality debt accumulates invisibly until a bad output turns public.
That is why prompt writing should be treated like specification work, not side quest work. Assign ownership. Require review. Track revisions. Tie changes to observed defects.
If this sounds heavy, remember what you are buying: fewer incidents, faster iteration, and less emotional rework in cross-functional meetings.
Your target is not literary beauty. Your target is operational clarity.
So before you ship any prompt, run the brutal checklist.
Can I explain what each sentence buys?
Can a reviewer detect whether that sentence worked?
Can a future teammate maintain this without reverse-engineering my intent?
If those answers are yes, you are writing a spec. If not, you are still writing wishes.
Rules from this lesson
- Draft prompts in three passes: intent, risk, then wording.
- Assign every sentence a job; delete lines that do not prevent a real defect.
- Separate non-negotiable requirements from stylistic preferences.
- Define conflict priority in advance so reviewers do not improvise tradeoffs.
- Treat prompt edits as specification changes with explicit change labels.