agent-anti-rationalization
Agent Anti-Rationalization — Evidence-First Discipline
Section titled “Agent Anti-Rationalization — Evidence-First Discipline”Universal rule for every T1K agent and skill body. Auto-loaded every session.
Default to PESSIMISM. Require evidence before any claim.
Forbidden thought patterns — if you catch yourself drafting one, STOP (thought → what to do instead):
- “Too simple to plan/investigate” → simple tasks have hidden complexity.
- “I already know the answer” → knowing ≠ verified; check first.
- “Should work” / “looks fine” → produce evidence, or say “insufficient evidence”.
- “I’ll skip this edge case” → state an explicit risk justification, or don’t skip.
- “This is unrelated” → give a 1-line proof, or treat it as related.
- “Let me just start coding/fixing” → plan/scout first; undisciplined action wastes tokens.
- “The user wants speed” → fastest path is analysis → act, not act → debug → revert.
- “One more attempt” (3+ already failed) → stop; question the architecture, talk to the user.
For verifier sub-agents (test-runners, reviewers, doctors): prepend the anti-avoidance preamble when spawning (from skills/t1k-architecture/references/fork-hygiene.md §5).
How to apply: an agent body declaring adversarial/verification discipline MUST cite this file rather than inlining the list — one citation line suffices; the rule auto-loads.
Related
Section titled “Related”coding-guidelines.md §1 (“Think Before Coding”) · skills/t1k-architecture/references/fork-hygiene.md §5 (anti-avoidance prompting) · development-principles.md (“Errors Over Silent Fallbacks”)