Prompting the current models

A model release is a prompt-review trigger, not just a changelog entry. model: opus in agent frontmatter silently resolves to whatever “opus” means today — so a prompt tuned against the previous model rides along and can quietly get worse.

Two of Opus 5’s behavioural shifts run opposite to Opus 4.8, which is what makes them easy to miss: the instructions written to correct 4.8 now push in the wrong direction.

Behaviour	Opus 4.8	Opus 5	What to do
Subagent delegation	Under-reached; needed nudging	Reaches freely	Delete “use subagents” nudges. Cap spawn counts explicitly instead
Self-verification	Had to be told to verify	Verifies unasked	Delete “double-check your work”. Removing it cuts over-verification with no loss of rigour
Task scope	Stayed put	Can widen the task, or apply its own judgment about what the task should be, without saying so	Keep explicit scope fences. They carry real weight now
Response length	—	Longer prose, and longer files written to disk	Ask for concision in words. Lowering `effort` moves thinking, not prose
Self-correction	—	Narrates its own earlier mistakes at length	Scope corrections to the ones that change what the reader does next

Three consequences, concretely

1. A severity filter in a reviewer is now a bug.

“Report only material findings” and “skip subjective style preferences” read as instructions to withhold. The model obeys literally: it investigates just as thoroughly, finds the defect, and then declines to report it as below the bar. Precision rises, recall falls, and real bugs disappear without a trace.

The fix is to move the filter. The reviewer reports every plausible finding in named categories with a severity and a confidence; the orchestrator filters. That filtering step already existed in the loop — it just was not being used as the filter.

Note the shape of the instruction: it is not “report absolutely everything”, which invites aesthetic noise. It is report everything within these categories, never omit for low severity or confidence, exclude pure aesthetics unless they break a stated rule, and consolidate same-root-cause findings. High recall without paying for taste.

2. “Don’t fix it yourself” needs a sibling.

An Opus orchestrator delegates readily — that is the role working as intended — but is also prone to re-deriving what the subagent already reported, paying twice at the expensive tier. So: if you delegate, commit to the delegation. Read the diff to verify the work, not to repeat it.

3. “Report concisely” is load-bearing.

Both Opus 5 and Sonnet 5 default to longer output than their predecessors, and the entire value of subagent isolation is that only a short report crosses back. Treat the concision instruction in the worker prompts as a live constraint, not boilerplate.

Effort

Opus 5 supports low through max. Start at high for review gates and sweep down — low and medium are unusually strong on this model, and an effort level inherited from a previous model is rarely the right one. max can overthink; do not reach for it reflexively.

Sonnet 5 in the worker seats

It follows instructions more literally. Holdover directives now apply at face value — re-read them rather than assuming they are inert.
It is more agentic: it reaches for tools and verifies on its own.
It gives better in-progress updates by default, so any “summarise progress every N steps” scaffolding can go.
Its effort defaults to high, and medium is a real cost lever for the orchestrator tier. This setup uses medium for orchestration and high for implementation — but a skill’s effort: applies for the current turn only, so the orchestrator holds medium only while its delegations stay in the foreground. Set the session’s effort too. See Gotchas.

Tokenization

Newer models produce meaningfully more tokens for the same text than older ones — roughly a third more is the figure quoted for the current generation, but the exact difference depends on the content. A cost or context baseline measured on an older model under-predicts. Measure rather than extrapolate, and do not treat a mixed-generation fleet as directly comparable per token.