Comparison of Anthropic's Prompting Guidelines for Claude Models

Comparison of Anthropic's Prompting Guidelines for Claude Models

Claude models respond similarly to most instructions, but there are a few key differences to be aware of. Keep these in mind when using the same prompt, skill, or agent with different models.

Category Fable 5 / Mythos 5 Opus 5 Sonnet 5 Opus 4.8 Summary
Effort high default. xhigh only for capability-critical work; low and medium are genuinely fine — they often beat xhigh on older models. high default. Use low/medium liberally as your primary cost lever; step to xhigh only for demanding agentic work. high default. Raise to xhigh for the hardest coding and agentic tasks. Start at xhigh for coding and agentic work; never below high for anything intelligence-sensitive. The arrow reverses. Carrying 4.8's always-xhigh habit into Opus 5 or Fable 5 burns tokens for no gain; carrying Opus 5's low is fine back to 4.8 under-thinks.
Thinking Adaptive only. No budgets, summarized output only. On by default. Can be disabled only at high effort or below. On by default — changed from 4.6. Turn off with thinking:{type:"disabled"}. Off unless you explicitly set {type:"adaptive"}. A prompt written for 4.8 assumes no thinking. The same request on Sonnet 5 or Opus 5 now spends max_tokens on thinking and can truncate mid-answer.
Verbosity One short brevity rule is enough — it replaces enumerating every pattern you dislike. Prompt for length explicitly. Lowering effort cuts thinking, not the visible answer. Length self-calibrates to task complexity. Tune only if your product needs a fixed style. Length self-calibrates. Tune only if your product needs a fixed style. On Opus 5, dropping effort to get shorter replies is the wrong knob. On 4.8 and Sonnet 5 it partly works, which is why the habit survives.
Verify-your-work Add it. Make self-verification explicit and run fresh-context verifier subagents on an interval. Delete it. Explicit verification triggers over-verification. Drop "double-check your answer" too. not flagged not flagged A flat contradiction. The long-run prompt that makes Fable 5 reliable makes Opus 5 pay twice for the same work.
Subagents Delegates readily — so use them more: async, long-lived, running in parallel. Delegates readily — so cap it. Never delegate verification of its own work; keep spawn counts low. not flagged Spawns too few. Prompt it to fan out across files and items. Three different directions. 4.8 needs encouragement, Opus 5 needs a leash, Fable 5 wants more delegation than either.
Prompt detail Strip it back. Skills tuned for older models are too prescriptive and actively degrade output. Remove legacy verification scaffolding; otherwise 4.8 prompts run as-is. Sonnet 4.6 prompts work out of the box. Prescriptive prompting is how you steer it. Fable 5 is the one model where your most carefully tuned legacy prompt is the liability. More specific stops being better.
Instruction style Give intent and the reason behind the ask. Short instructions steer broadly. Constrain scope explicitly on narrow tasks — it expands scope on its own. Literal. Won't generalize an instruction from one item to the next. State the scope. Literal. Same as Sonnet 5. Sonnet 5 and 4.8 need every case spelled out. Fable 5 penalizes that same enumeration.
Progress scaffolding Replace it with a send_to_user tool plus elicitation language, and add a re-grounding rule for final summaries. Remove it, then tune narration down — Opus 5 over-narrates. Remove it. Updates are already good. Remove it. Updates are already good. Only Fable 5 wants new machinery here. Only Opus 5 needs the volume turned down after you delete the old scaffolding.
Reasoning rules Never ask it to echo or explain its reasoning — that trips the reasoning_extraction refusal and elevated fallbacks. Never tell it "don't think" or "don't reason" — that increases <thinking> tag leakage when thinking is off. Opposite failure modes, same root cause. Both resolve to: write no rules about reasoning at all.
Sampling parameters Not accepted. Not accepted. 400 error on temperature, top_p, top_k — new for Sonnet-class models. Not accepted. Anything using temperature for stylistic or design variety has to move to asking the model for several directions up front.
Frontend defaults Settles into a house style; generic negatives just move it to a different fixed palette. A documented house style: cream #F4F1EA, serif display, terracotta accent. "Don't use cream" fails on both. Give a concrete spec, or ask for four directions and pick one.
Code review recall Higher bug-finding recall, but cyber and bio safety classifiers can decline findings outright. Ask for everything; filter in a second pass. Ask for everything; filter in a second pass. Ask for everything; filter in a second pass. The rare axis where all four agree — and it still breaks you. "Only high-severity" is now obeyed literally, and measured recall collapses.
Watch for Runs for minutes to hours, so fix client timeouts. Ground progress claims in tool results. State boundaries on unrequested actions. Give it a memory file. Don't show it a context countdown. Files it writes run long — calibrate document length. With thinking off, tool calls can leak into the response as plain text. Tool use is up versus 4.6 but drops with thinking off, so nudge explicitly. The new tokenizer produces roughly 30% more tokens. Prefers reasoning over tool calls. Raise effort to get more tool use. Same knob, opposite ends: Sonnet 5 and 4.8 need agentic behavior pushed up, Fable 5 and Opus 5 need it reined in.

Sources