Source Intelligence
Follow Discord

DisclaimerUnofficial, and not affiliated with Anthropic. Nearly all of this is read straight out of what ships: npm bundles, captured prompts, published docs. Anthropic's own notes go in verbatim, marked as theirs. The rest is my reading, and every entry carries the strings behind it. If one looks wrong, vote it down and say why.

← All of v2.1.86 Home All releases ←olderv2.1.85 v2.1.87newer→
Claude Code v2.1.86

Agent System Prompt Refinements

Agent system prompts have been refined with more specific guidance:

  • New "don't gold-plate" phrasing: "Complete the task fully—don't gold-plate, but don't cut corners"
  • More emphatic $TMPDIR guidance: "TMPDIR is automatically set to the correct sandbox-writable directory" (removed fallback variable mention)
  • New guidance on failure handling: "diagnose why before switching tactics—read the error, check your assumptions, try a focused fix"
  • Fork workers now use a "Your directive:" prefix with a "fork-boilerplate" key instead of embedded boilerplate text
  • Summary format now strictly requires <analysis> and <summary> blocks with tool-call prohibition: "CRITICAL: Respond with TEXT ONLY. Do NOT call any tools"
Evidence

Agent prompt changes (search for "Your directive:" and "CRITICAL: Respond with TEXT ONLY")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

See this entry in the whole of v2.1.86 →