The Prompting Playbook Just Changed. Both Claude and GPT Now Punish Vague Prompts.

If your AI outputs have gotten flatter in the past few weeks, your prompts are probably the culprit. Both Claude 4.7 and GPT-5.5 now penalize vague prompting — but in different ways, which means the single approach that worked across both platforms no longer does. The Neuron reported on May 1 that the gap between a specific prompt and a broad one is now noticeable enough to affect real work product.

What changed

Claude 4.7 rewards surgical specificity. Broad instructions that produced decent output in Claude 3 now produce noticeably degraded results — the model is less forgiving of missing context and underspecified scope. GPT-5.5 takes a different approach: it performs best when you state what success looks like, not just what you want done. Same vague prompt, two different failure modes.

OpenAI published a structured prompt template this week with six elements: role, goal, success criteria, constraints, output format, and stop rules. That last one matters — without a stop rule, GPT-5.5 tends to keep elaborating well past what you needed. (The Neuron, May 1.)

TLDR AI also reported (Apr 29) that Anthropic's Opus 4.7 tokenizer raised effective costs 12–27% on longer prompts despite unchanged list pricing. Tighter prompts are not just better — they're now cheaper.

What to do differently Monday morning

You need two moves. First, scope your request before sending it: what specifically do you need, what should be excluded, and what does a good result look like? Second, add a success criterion — one sentence describing what the output must do, not just what it must contain.

For Claude, specificity is the variable that drives quality. For GPT, outcome-framing is what activates the model's best reasoning. Both reward the same discipline: knowing what you need before you ask for it. Ethan Mollick's GPT-5.5 review (Apr 23, still generating discussion) made a related point — structured analytical tasks work well, but nuanced or creative work still needs heavy human oversight. Knowing the model's current strengths is part of prompting well.

The campaign brief that started producing noise

Priya leads content at a 25-person B2B tech company. She had been using the same prompt across both tools: "Write a LinkedIn post about our new product launch." Flat output. Generic framing. Nothing that matched the tone her company had spent six months building.

After reading about the model changes, she rebuilt the prompt using the six-element structure:

Role: You are a B2B content writer familiar with enterprise software buyers.
Goal: Write a LinkedIn post announcing our product's integration with Salesforce.
Success criteria: The post should make a VP of Sales at a 200-person company stop scrolling.
Constraints: No jargon. Under 150 words. No exclamation points.
Output format: Three short paragraphs, no headers.
Stop rules: Do not add a call to action at the end.

The output on both platforms improved significantly. More importantly, the exercise forced her to decide what a good post actually looked like before generating one — a step she had been skipping.

The RFP summaries that got too long

James runs vendor operations at a 60-person professional services firm. He had been using AI to draft RFP summaries — a task that worked reliably three months ago but started producing bloated, unfocused outputs after recent model updates.

The fix was adding constraints he had never needed to specify before. Not just "summarize this RFP" but: "Summarize this RFP in three bullet points. Focus only on scope, timeline, and budget. Do not include background on the vendor or their other products."

For Claude 4.7, the specificity fixed the problem immediately. He ran the same prompt in GPT-5.5 and found it still needed one more piece: a success criterion. He added "A project manager should be able to brief their team from this summary without reading the original document." That sentence changed the output quality considerably more than any structural adjustment had.

The one-sentence rule

A prompt that worked last quarter is not a prompt that works today. Both models now surface the quality of your thinking before they surface the quality of their output — if you haven't specified what good looks like, neither can they. Before you send any request longer than two sentences, write one sentence that describes what a successful result must do. That sentence alone closes most of the gap.