Most Claude Code setups in Australian businesses were written for an older model. The CLAUDE.md files, the effort settings and the habit of reaching for Plan Mode on every task all date from a time when the model needed more hand-holding. Claude Opus 5.5 needs less of it, and carrying the old scaffolding forward now costs money and slows teams down.
This post is about what to change, based on practical guidance shared in September 2026 by builders working with Opus 5.5 day to day. The short version: the better the model gets, the less prompt engineering you should be doing.
How should you prompt Claude Opus 5.5 differently?
Prompt Claude Opus 5.5 with less scaffolding than you used for earlier models. Leave effort on the default medium setting, remove reasoning instructions such as "think carefully" from CLAUDE.md, add one explicit rule telling Claude to keep working until it genuinely needs input, and expect to use Plan Mode less often as the model gets better at judging what needs approval.
Each of those four changes removes something rather than adding it. That is the pattern worth noticing. Teams that respond to a new model by writing longer instructions tend to get slower, more expensive sessions without better results.
Leave effort on medium
Effort tier names do not carry their meaning across model generations. According to third-party benchmarking cited in the guidance, Opus 5.5 on medium effort matches the quality Opus 5 delivered on max effort, at roughly a quarter of the cost. A developer who bumps effort to max out of habit is paying for depth the model would have found anyway.
There is one practical catch. Switching effort tiers in the middle of a session breaks the context cache, so the next turn pays full price to re-read everything. Decide the effort level once at the start of a session and leave it. If a task turns out to need more depth, start a fresh session for it rather than flipping the setting halfway through. Our breakdown of what one Claude Code task really costs on Opus 5.5 shows why cache hits matter so much to the final bill.
Prune reasoning instructions from CLAUDE.md
Lines like "think step by step", "reason carefully before acting" or "double-check your work" made sense for earlier models. Opus 5.5 sets its own reasoning depth, and those instructions now add tokens to every turn without changing the outcome much. In some cases they push the model into longer deliberation on simple tasks.
Anthropic's own /claude-api prompt-audit workflow is built to flag exactly these anti-patterns. Run it against any CLAUDE.md written before Opus 5.5. We covered a manual approach in our CLAUDE.md audit checklist, and the same thinking applied to Opus 5 in five tuning patterns for Claude Code. Opus 5.5 pushes the same direction further.
Things worth keeping in CLAUDE.md are facts the model cannot infer: build commands, where tests live, naming conventions, which directories are off limits, and who approves what. Things worth cutting are instructions about how to think.
Add one explicit keep-going rule
This is the one change that adds a line rather than removing one. On long-running tasks, Opus 5.5 sometimes reports progress and then stops, waiting for a reply nobody intended to give. For a developer watching the terminal it is a minor irritation. For an unattended job it means the work simply does not get done.
A short house rule in CLAUDE.md closes most of that gap:
Keep going when the next step does not need input from a person.
Put any status update in the same message as the next action, not on its own.
Stop only when you genuinely cannot continue, or before anything destructive or irreversible.
The last point matters. The rule is not "never stop". It is "stop for the right reasons". Deleting data, pushing to production, sending messages and spending money should still pause for approval.
Expect Plan Mode to matter less
Plan Mode was built for models that needed a structured pause before acting. Anthropic's Claude Code team has floated removing it entirely as models improve. Whether or not that happens, the direction is clear: Opus 5.5 is better at telling which actions need sign-off and which do not, so heavyweight planning steps earn less than they used to.
For most routine tasks, a clear request and a well-kept CLAUDE.md now do the job Plan Mode used to do. Keep it for large refactors, unfamiliar codebases and anything touching production infrastructure, where a written plan is useful for the human reviewer as much as for the model.
What changes, at a glance
| Habit | Older models | Opus 5.5 |
|---|---|---|
| Effort setting | Max for anything important | Medium by default, set once per session |
| Reasoning lines in CLAUDE.md | Helped quality | Mostly cost tokens, prune them |
| Long unattended tasks | Rarely attempted | Add an explicit keep-going rule |
| Plan Mode | Used on most tasks | Reserve for large or risky changes |
| CLAUDE.md length | Grew over time | Shorter, facts only |
What this is worth to an Australian team
An illustrative example: a Sydney engineering team of 12 developers spending $45,000 a year on Claude Code usage, much of it on max effort out of habit. If moving routine work to medium effort and trimming CLAUDE.md cut usage cost by even a third, that is roughly $15,000 a year back, before counting the time saved on sessions that no longer stall waiting for input. The real figure depends on your mix of tasks, so measure a fortnight before and after rather than trusting the estimate.
The time saving often matters more than the dollars. A developer who no longer has to nudge Claude every few minutes on a long task gets their afternoon back.
What not to conclude
Doing less prompt engineering is not the same as doing none. Opus 5.5 still needs to know your build commands, your conventions and your boundaries. It still needs approval gates on destructive actions, and regulated teams working under APRA CPS 234 or the Privacy Act still need logging and review on what the agent touched. The point is to remove instructions about how the model should think, not instructions about your business.
It is also worth re-running your own evaluations after each change. Third-party benchmarks are a guide, not a guarantee for your codebase. See what changed in Opus 5.5 for the broader release picture.
Where to start this week
Pick one active repository. Read its CLAUDE.md and delete every line that tells Claude how to think. Add the keep-going rule. Set effort to medium and leave it there for a week. Compare the bill and the number of times someone had to nudge a session along.
If you are rolling Claude Code out across several teams and want a second set of eyes on the setup, our Claude services cover exactly this kind of audit. Or book a short call and we will talk through what your current CLAUDE.md files are costing you.
