1300 166 138 Tell us what you need
News / Claude Code

Claude Code says less now and developers want proof

By Monika Tantau • Published 3 months ago
Claude Code says less now and developers want proof

Key takeaways

• Anthropic added a built in Concise output style to Claude Code in version 2.1.237
• Concise leads with the result and skips preamble, while error reports and security warnings stay complete
• A Hacker News thread with more than 180 points asked whether effort levels had been quietly reduced
• Anthropic said the effort you select is the effort you get, and offered credits for genuine regressions

Most Read

Anthropic has added a setting to Claude Code that tells the model to stop narrating its own work. The Concise output style, which arrived in version 2.1.237 on 20 August, makes Claude lead with the result and drop the preamble that developers had been complaining about for months.

The timing is awkward. Within days, a thread on Hacker News accused Anthropic of quietly dialling down how hard Claude works on each request, and a member of the Claude Code team had to answer it in public. From the outside, brevity and degradation look identical.

What the Concise style actually does

Concise is one of five built in output styles that now ship with Claude Code, alongside Default, Proactive, Explanatory and Learning. Output styles work by modifying the system prompt, so the change applies to every response in a session rather than one request at a time.

The style tells Claude to lead with the result, skip preamble and narration, and keep responses short by default. The engineering work itself stays as thorough as the Default style, according to Anthropic’s documentation, and Claude keeps the complete content of error reports, security warnings and confirmations for destructive actions. Ask for an explanation and it still gives one in full.

Turning it on takes a single setting. Developers pick it from the config screen in the terminal, or set the outputStyle field to Concise in a settings file. Because output styles are read once when a session starts, the change only takes effect in a new session.

There is a billing side to this as well. Output styles move token usage in both directions: Explanatory and Learning produce longer answers by design, and Concise does the opposite by keeping responses short. For a team on usage based pricing, less narration is not only easier to read, it is fewer tokens to pay for on every turn.

Why the narration became a problem

Claude Code’s default behaviour is to explain what it is about to do, restate the task, then describe each step as it goes. On a small change that means reading several paragraphs to find the one line naming the file that moved. Developers running the tool for hours a day were paying that reading cost on every single turn.

The complaint had already produced workarounds. Developers wrote their own output styles, pushed instructions into project memory files, and in at least one case built a local hook that rewrites Claude Code’s output before it reaches the terminal. Anthropic has framed Concise as an interim measure while it works on a deeper fix.

The effort question that followed

The effort question that followed

On Saturday a Hacker News thread arguing that Anthropic was testing reduced effort levels in Claude Code passed 180 points and 160 comments. Users reported that the numerical effort indicator was showing lower values than the level they had selected, and read that as evidence the model was being throttled.

Thariq, from the Claude Code team, answered in the thread. Anthropic sometimes tests API serving configurations before rolling them out, he wrote, and one running at the time mapped the numerical effort value differently. His summary was blunt: the effort you selected is the effort you are getting. He also offered credits to anyone reporting a genuine regression.

The replies did not settle it. One developer described a configuration file update that took 43 minutes on Opus 5 and under two minutes on an earlier version. Others pointed at the gap between rising benchmark scores and their own sense that the tool had become less useful. None of that is measurement, which is exactly the problem.

Why Australian businesses should pay attention

For an Australian team this is a supplier management question before it is a technical one. A coding assistant that changes its output shape and its effort scale between releases is a dependency whose behaviour moves without a purchase order. Most software vendors cannot alter a product this fast, and most procurement processes were never built for one that can.

The practical answer is to hold your own baseline. Pick a handful of tasks the team runs often, record what good output looks like, and rerun them after each significant update. That takes an afternoon to set up and it turns an argument about impressions into a number. It is ordinary AI governance work, and it is cheap next to the cost of finding out late.

Cost gives the change a second dimension. A firm putting a coding assistant in front of a dozen developers is buying output tokens at volume, and narration nobody reads is spend that produces nothing. Concise turns a readability complaint into a line item worth tracking beside the quality baseline.

Tool choice is affected too. Melbourne teams weighing Claude Code and Codex now have to price in how each vendor communicates change, not just how each one scores on a benchmark. The same question applies across the wider field of coding assistants.

What to watch from here

Concise is a real improvement for anyone reading Claude Code output all day, and the guardrails are sensible. Error reports and confirmations for destructive actions stay complete, which is precisely where brevity would do damage.

The larger issue is disclosure. Testing serving configurations on live users is normal engineering practice, but it means the tool a developer used on Friday is not guaranteed to be the tool they open on Monday. Publishing which experiments are running would cost Anthropic very little and would remove most of the suspicion in one step.

For Australian businesses building on these tools, the lesson is smaller and more useful than the argument around it suggests. Measure your own outputs, keep the record, and treat a vendor’s description of its own product as one input rather than the answer.

Scroll to Top