Michael Chihlas
50215b9110
All checks were successful
Mirror to GitHub / mirror (push) Successful in 10s
The system prompt had a "Complete example of a correct first response" section with a specific Outlook/WiFi/jsmith scenario plus literal JSON payloads in [QUESTIONS], [ACTIONS], [SUGGEST_FIX], and [PROMOTE] markers. The model was emitting those literal strings (the same WiFi/laptop questions, the same "Clear cached credentials" suggested fix, the same "OWA login confirmed for jsmith" promote) on EVERY unrelated chat — making the task lane look like it was leaking previous- session data when in fact the AI was just reciting the prompt examples. Replaced literal example content with `<placeholder>` schemas. Added an explicit ANTI-PARROT RULE in the FINAL REMINDER section calling out that the angle-bracket placeholders show SHAPE, not CONTENT, with concrete examples of the failure mode (printer ticket → don't ask about Outlook; user not named jsmith → don't name jsmith). Same scrub applied to the FORK section's "Outlook AND Teams dropping" and the worked fork-flow example. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>