Prompting Claude 5 models is mostly a subtraction exercise. Claude Opus 5 and Claude Fable 5 follow instructions well enough that rules written to protect against older models now cost quality instead of adding it. Anthropic removed more than 80 percent of Claude Code's system prompt for these models with no measurable loss on its coding evaluations. Start by deleting, then tune verbosity and effort.
What actually changed with the Claude 5 generation?
The guardrails became the problem. When a rule is written to prevent a specific failure, it freezes one answer to a question the model can now answer better by reading the situation.
Anthropic's own example is comment density. Older versions of Claude Code carried a strict rule to write essentially no comments and never write multi-line comment blocks. That rule was correct on balance for older models, and wrong for a real subset of cases: some code genuinely needs a block comment, and some users want documentation. The replacement is a judgment instruction rather than a prohibition, telling the model to write code that reads like the code around it, matching its comment density, naming, and idiom.
There is a second cost that is easier to miss. Instructions accumulate in layers, written by different people who cannot see each other's work: a system prompt, a project instruction file, a skill, and your actual request. Anthropic found conflicting messages inside single requests in its own internal transcripts, one layer asking for documentation while another forbade comments. The model usually resolves the conflict correctly, but it has to spend reasoning on the conflict before it can spend any on your problem. Stale rules do not just waste context. They make the model think about you instead of the task.
What should you delete from a prompt written for an older model?
Work through your prompts, project files, and skills and sort every instruction into three buckets. This is the audit, in the order that finds the most in the least time.
1. Delete anything that instructs verification. This is the highest-yield category and the most counterintuitive, because these lines feel responsible. Claude Opus 5 verifies its own work without being told. Instructions like "include a final verification step" or "double-check your answer before responding" compound with behavior the model already has, so they burn tokens without improving the result. Anthropic's guidance is to remove them outright. The same goes for harness scaffolding that bolts a separate verification step onto every task.
2. Delete rules that forbid a behavior you have not actually seen from this model. Most defensive rules were copied from advice about a model you are no longer using. The test is simple: can you point at a session where this model did the thing your rule forbids? If not, the rule is protecting you from a ghost, and it is competing with your real instructions for the model's attention.
3. Delete examples that were teaching tool usage. This one reverses a rule that was true for years. On Claude 5 models, examples constrain the model to the space the examples describe rather than showing it what good looks like. Where you were using examples to explain a tool or a script, put the explanation into the design of the thing itself: expressive parameter names, a clear list of allowed values, a one-line description on the tool rather than a paragraph in the prompt.
4. Move anything rarely needed out of the always-loaded file. If a block of instructions matters on one task in ten, it should not be paid for on all ten. Put it in a separate file or a skill that gets loaded when relevant. Anthropic calls this progressive disclosure, and its own example is moving code review and verification instructions out of the system prompt and into skills that get called when needed.
5. Keep the gotchas. What survives the cut is the knowledge the model cannot infer by looking at your repository: the conventions that are not visible in the code, the thing that will break if touched, the opinions specific to your team. Everything a model can learn by reading your file system is wasted budget. What belongs in a project instruction file, and what does not, is covered in CLAUDE.md and AGENTS.md.
A shortcut exists for Claude Code users: the /doctor command is Anthropic's own tool for rightsizing skills and instruction files, and it encodes this guidance.
Why is Claude Opus 5 so verbose, and how do you shorten it?
Ask for brevity in words. Opus 5's default responses are longer than earlier Opus models, both in conversation and in the files it writes to disk, and this is the single most common complaint after an upgrade.
The mistake is to reach for effort. Effort controls how much the model thinks, not how much it says, so turning it down can reduce reasoning without reliably shortening the answer. What works is a short, direct instruction: tell it to keep responses focused, keep caveats brief, spend most of the response on the main answer, and give a high-level summary unless you ask for depth. If your system prompt is long, repeat a one-line reminder near the end of it.
Written deliverables need their own instruction, separate from conversational length. If Claude is producing reports or documents, tell it to match the length to what the task needs and to skip filler sections, redundant summaries, and boilerplate.
Narration is a third, separate dial. Opus 5 announces what it is about to do more than earlier models. Describing the cadence you want works better than forbidding narration: one sentence before the first tool call, brief updates only when something important turns up or the direction changes, and a final message that leads with the outcome. Positive descriptions of the style you want beat instructions about what not to do.
What are effort levels and which one should you use?
Effort is the API parameter that trades intelligence against latency and cost, with four settings: low, medium, high, and xhigh. Both models default to high.
For Opus 5, Anthropic recommends starting at the default and then testing lower settings on your own work, using low and medium liberally wherever quality holds, and reserving xhigh for demanding agentic tasks. For Fable 5, high is the everyday setting, xhigh is for the hardest work, and low or medium handle routine tasks well enough that Anthropic claims they often beat the previous generation at its maximum.
The instruction that applies to everyone: if you carried an effort default over from an older model, test it again. The right level is model-specific and your old answer is stale.
One thing effort does not fix is scope. At higher effort on a routine task, these models deliberate more than the job needs and can start tidying code you did not ask them to touch. The fix is a scope instruction, not a lower effort setting, because you usually want the reasoning and not the cleanup.
If you use Claude Code rather than the API, the think and ultrathink keywords are the chat-level surface of the same idea. Our guide to extended thinking mode covers when each level pays off.
Leave thinking on. Opus 5 allows disabling it only at high effort or below, and Anthropic reports that thinking enabled at low effort outperforms thinking disabled at similar cost. With thinking off, two glitches occasionally appear: the model writes a tool call into its visible text instead of actually calling the tool, and internal tags leak into the response.
Should you still ask Claude to check its own work?
Not on a single task. Yes on a long one, and the difference is who does the checking.
On a normal task, the model already verifies, so asking again is pure cost. On a long autonomous run, the useful pattern is a separate verifier with a fresh context, checking the work against the specification at an interval, which Anthropic reports outperforms asking the model to critique itself. The reason is the same one behind our advice to review in a clean session rather than the one that wrote the code: a model that has been working for hours is judging its own output from inside the context that produced it.
Long runs need one more instruction, and it is the most important one on this page for anyone leaving an agent working unattended. Tell the model to check each progress claim against an actual result before reporting it, and to say plainly when something is unverified. Anthropic reports this nearly eliminated fabricated status reports in its testing, including on tasks designed to provoke them.
Code review deserves its own warning. Telling Opus 5 to report only high-severity issues makes it report less, because it follows the instruction literally. Ask for everything and filter in a second pass.
When should Claude use subagents?
When the subtask is genuinely independent and big enough that giving it its own context is cheaper than sharing yours.
Both Claude 5 models delegate more readily than their predecessors, and Anthropic steers them in opposite directions. For Opus 5 it recommends capping delegation, because spawning agents for small tasks multiplies cost and time. For Fable 5 it recommends delegating freely, keeping subagents long-lived across subtasks, and working while they run instead of waiting for each to return.
That is a price and horizon difference rather than a contradiction. The practical consequence for anyone paying by the token: an agent that spawns agents multiplies your spend without asking, and the multiplier is invisible until the bill arrives.
How do you work with Fable 5 on long tasks?
Expect it to run for a long time, and plan around that rather than against it. Individual requests on hard tasks take many minutes at higher effort, and autonomous runs can go for hours. Anthropic tells integrators to adjust timeouts and progress indicators before migrating, and it is worth pointing Fable 5 at genuinely hard problems rather than testing it on easy ones, since easy tasks hide what it is for.
Three habits matter more here than on shorter work:
- Say when to stop and ask you. Define the cases that genuinely need you, such as a destructive action, a real scope change, or information only you have, and let it proceed on everything else. Without that, it either interrupts too often or, occasionally deep into a session, stops and states an intention without acting on it. A "continue" restarts it.
- Give it a place to write lessons down. These models perform well when they can record what they learned in one run and read it in the next: one lesson per file, a summary line at the top, and pruning of anything wrong or already recorded elsewhere.
- Give the reason, not just the request. Explaining what the work is for and who it is for lets the model connect the task to relevant context instead of inferring your intent.
How do you find what you forgot to tell it?
Ask before you build, because a faster model implements the wrong specification just as completely as the right one.
The best-known method for this comes from Anthropic's field guide for Fable, which sorts the problem by what each side knows. The hardest category is not the gaps you can see, it is the things so obvious to you that you never thought to say them, plus the ones neither of you anticipated. Four techniques target them:
- A blind spot pass. Ask the model what you are failing to consider in an area you do not know well. It is the only one of these aimed at gaps neither of you can name.
- An interview. Have it ask you one question at a time about anything ambiguous, prioritizing questions whose answer would change the architecture. That prioritization is what separates a useful interview from a trivia quiz.
- Prototypes to react to. Ask for several deliberately different directions and respond to them. Reacting surfaces preferences you could not have written down.
- A quiz, before you merge. Have the model test your understanding of the change it just made. This is the one worth adopting even if you skip the rest, because it catches the real risk in fast agentic work: shipping code nobody on your team understands.
For the surrounding workflow, see how to use Claude Code properly, and for managing the context these sessions consume, scaling AI development. Launch details for each model are on the Claude Opus 5 and Claude Fable 5 pages.
FAQ
Do I need to rewrite my prompts for Claude Opus 5?
Not rewrite. Audit and cut. Anthropic says Opus 5 works well out of the box on prompts written for Opus 4.8, so nothing breaks if you change nothing. The gains come from deleting verification instructions, defensive rules for failures you have not seen, and examples that were teaching tool usage.
Why does Claude Opus 5 write so much?
It is the default behavior of the model, not a setting you have wrong. Lowering effort will not reliably shorten it, because effort governs thinking rather than output length. A short explicit instruction asking for concise responses is what works, applied separately to conversation and to any documents it writes.
What effort level should I use with Claude?
Start at the default of high, then test low and medium on your own tasks and keep the lowest setting where quality holds. Reserve xhigh for the hardest agentic and coding work. If you set an effort default for an earlier model, re-test it rather than carrying it over.
Should I turn thinking off to save money?
No. Use a lower effort level instead. Anthropic reports that thinking enabled at low effort beats thinking disabled at comparable cost, and disabling thinking introduces occasional output glitches, including tool calls that appear as text and never actually run.
Is Claude Fable 5 or Opus 5 better for coding?
Opus 5 for most everyday coding, on price. Fable 5 for long-horizon work measured in hours or days, and for genuinely ambiguous problems where you want the model to work out the next steps. Fable 5 costs twice as much per token, and Anthropic's own benchmarks put Opus 5 within a fraction of it on several coding evaluations.