Claude Opus 5.5 Prompting: Speak Constraints First
A better way to use a faster writing model: say what cannot change before asking it to polish.
September 28, 2026
Claude Opus 5.5's new prompting guide is worth reading for one reason: it treats speed, effort, and completion as separate controls. That is a better way to use any strong writing model. Do not just describe the task. Say what must stay true, what must not happen, and how you will know the work is done.[1]
Anthropic says Opus 5.5 generates output more than 30 percent faster than Opus 5 and costs 40 percent less on typical workloads. Its own documentation also says effort is the first setting to change when you are trading off intelligence, latency, and cost.[1][2] The attention is real. The prompting guide reached the Hacker News front page with 168 points and 178 comments when checked on September 28.[3]
The useful part of the new guide
Anthropic's guide is unusually blunt about a mistake that causes a lot of bad AI work: an agent reporting progress is not necessarily an agent that finished. On long tasks, Opus 5.5 can end a turn with a text update. The guide says an unattended loop should treat that as a report, keep a checklist of open parts, and only stop when the completion condition is met.[1]
That advice applies outside agent coding. If you ask a model to clean an email, it can produce a very polished paragraph and still drop the date, soften a boundary, or change who is making the ask. The answer may look complete because the prose is complete. Those are different things.
The guide also puts effort calibration ahead of clever wording. Start at medium, test against your own examples, and raise effort only when you have measured a gain. That is a healthier habit than treating every new model as a reason to write longer prompts.[1]
Speak the constraints before the details
A reliable prompt has a small contract inside it. Say the job in one sentence, then say the facts that cannot move. For an email, that might be the price, date, audience, and next step. For a customer update, it might be the promise you are not willing to make. For a meeting summary, it might be which points are decisions and which are still guesses.
Voice is good at this because people often say the guardrails naturally before they polish them. "Keep this short. Do not promise a refund. Mention Thursday, not Friday. Ask whether she wants a screenshot." Those little qualifiers tend to disappear when someone types a rushed one-line request.
A useful spoken prompt is not a monologue. It is a short instruction followed by a few non-negotiables:
- Turn these notes into a five-sentence customer reply.
- Keep the price at $69 per year.
- Do not claim the feature is live yet.
- End by asking whether they want access to the beta.
- If a fact is missing, leave a bracketed question instead of guessing.
That last line is the one most people skip. It prevents the model from filling a gap with a clean sentence that nobody approved.
Do not confuse a good voice with a good result
Anthropic says its newer model communicates more naturally and puts important information earlier.[2] That can make review easier. It can also make an error easier to miss. A tidy answer can sound more certain than the notes it came from.
The practical review is short. Compare the source and final version for the number, date, name, recipient, and action. Then read the first and last sentence. Those are where a model often changes the stance of a message. "Could I send an example?" and "Send an example" are not the same request.
For work that matters, give the model a named finish line. For example: "Stop only after you have preserved the facts above, included the exact next step, and listed anything you could not verify." The model may still make a mistake. But now the mistake is visible and easy to check.
Where DictaFlow fits
DictaFlow belongs before the model, not in place of review. Hold a hotkey, say the rough message and the constraints, then release to put the text where your cursor is. The useful part is getting the real draft out quickly while the context is fresh.
If you use AI cleanup afterward, keep it narrow. Fix transcription friction, punctuation, filler words, and obvious repeats. Do not hand it permission to invent a stronger promise, a calmer tone, or a made-up detail. That matters in customer support, sales, professional notes, and any workflow where a small wording change changes what you are committing to.
This is also why system-wide dictation matters. The constraints are not only written inside an AI chat box. They show up in an Outlook reply, a ticket, a CRM note, a Teams message, or a remote desktop field that does not like clipboard paste. DictaFlow's Citrix guide explains how it handles those tougher insertion paths.
The prompt test worth keeping
The most useful takeaway from the Opus 5.5 guide is not a magic sentence. It is the idea that every model run needs a completion condition. For writing, that condition is usually simple: the final text must preserve the facts, the boundary, and the next step.
Say those three things out loud before you ask for cleanup. Then check them once before you send. Faster models give you more chances to move quickly. They do not remove the need to decide what you mean.
Sources
[1] Prompting Claude Opus 5.5, Anthropic