4.6 KB of instruction cut Claude Code's answers from 524 words to 258

Clear Partner is one Markdown file with a frontmatter block, loaded as the communication layer of the system prompt ^[2]. The frontmatter sets `keep-coding-instructions: true` and `force-for-plugin: true` ^[3], so Claude Code's own coding instructions stay in place and only the reporting habits change. The instruction body runs about 4.6 KB ^[4].

What separates this from a brevity prompt is written into the instructions: "Economical in what the user reads, never in the work" ^[5], followed by a flat prohibition, "Never drop a warning, constraint, assumption, exact number, scope condition, or important tradeoff merely to be shorter" ^[6]. One carve-out survives for safety-critical warnings, which stay as a single short line with nothing added ^[21]. The author wrote that "Most brevity prompts quietly make the model shallower" ^[18].

Four runs per arm produced those averages, so 12 answers per arm and 24 recorded answers in total ^[8]. Each answer lost 266 words on average ^[9]. For that to transfer, two things have to hold on your workload: your questions have to resemble the three chosen here, and prose has to be most of what you read.

The token accounting is less flattering than the word count. At four characters per token, 4.6 KB is roughly 1,150 tokens riding in the system prompt on every request ^[10]; at 1.33 tokens per word, 266 words is roughly 350 tokens of output saved per answer ^[11]. Prompt caching makes the standing 1,150 cheap after the first call.

The suite is 12 cases, run with `claude plugin eval`, and the first six pin the basics ^[12]. Six cases at $1.54 and 342 seconds is about 26 cents and 57 seconds per case ^[14][16]. The post flags that these figures come from one run per arm ^[25].

Two of the three layers sit on documented surface: an output style and a statusline ^[1]. The third, Clear Transcript, redraws the transcript using what the post calls Claude Code's undocumented function hooks, and it is the only layer filed under `experimental/` ^[22][1]. Undocumented interfaces change in patch releases without a deprecation notice, and the eval suite grades final messages; nothing in it grades a hook signature.

The design rule behind the split is stated plainly: prompts for judgment, deterministic mechanisms for mechanics, and never solve the same problem in two layers ^[20]. The measurements cover length and whether six basics still pass. Depth is the claim the whole design rests on, and according to the post it is asserted from experience with naive brevity prompts, which drop the warning, the exact number, the assumption and the tradeoff ^[19].