Stronger Models Want Thinner Prompts

The strongest July 21 signal came from Digg surfacing Simon Willison’s annotated fireside chat with Anthropic’s Claude Code team. The headline detail was surprising but consistent with what stronger models have been hinting at for months: the Claude Code team reportedly cut its system prompt size by roughly 80%, and newer models performed better with fewer examples and fewer “don’t do this” constraints. At the same time, the team described a richer surrounding system: proactive Slack-based agent work, team memory, automated outer-layer review, and explicit permission patterns. The practical synthesis is that better models reward thinner prompt scaffolding when the rest of the runtime is well designed.

Many teams still respond to agent misses by adding more prompt text. That works until it does not. Long prompts become hard to reason about, expensive to maintain, and fragile when they mix policy, examples, workflow instructions, and edge cases in one blob. The July 21 shift is toward cleaner separation of concerns. Put stable process into skills or playbooks. Put preferences into memory. Put risk controls into permissions and verifiers. Leave the live prompt to frame the actual task. That makes the system easier to debug, cheaper to evolve, and more portable across model changes.

Refactor one bloated agent workflow into thinner prompt layers. Trigger: a prompt that has grown through repeated patching, exceptions, and copied examples. Context: separate the current instruction set into task framing, reusable procedure, house policy, user preference, and completion checks. Tools: move long procedures into skills or workflow files, stable preferences into memory, dangerous actions into explicit permission controls, and quality rules into review or eval checks. Verifier: compare the old and new setup on a small task set, measuring output quality, latency, cost, and failure clarity rather than relying on feel. Budget: cap how much prompt text survives in the live task and how many exceptions you allow before creating a reusable artifact. Artifacts: produce the slimmed prompt, the extracted skill or checklist, the memory rules, and the validation results. Stop condition: the workflow only ships when the leaner system matches or beats the old prompt while becoming easier to inspect and revise.

Take one agent prompt over 1,000 words and label every sentence as one of five things: task, procedure, policy, preference, or verifier. Anything that is not task framing is a candidate to move out of the prompt and into a more stable surface.