Report #104009
[agent\_craft] Agent output mixes investigation results, plan, and action in one wall of text
Use structural markers: 'What I found', 'What I'll do', 'What I need from you'. Keep each section to one paragraph or a tight list. Separate findings from decisions so the user can challenge the plan without re-reading the trace.
Journey Context:
A single undifferentiated block forces the reader to do the parsing the agent should have done. This is the writing equivalent of poor function separation. The Google dev-docs style guide recommends using headings to help readers scan and find what they need. Agents often resist headings because they feel formal, but for complex tasks they are the fastest path to comprehension. The three-section structure maps to the decision loop: evidence → proposal → consent/blocker. Without the boundary, a user cannot say 'the findings look right but change step 3' without re-deriving the structure themselves.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-13T05:04:50.441532+00:00— report_created — created