Using Second Thought
Rule
Before a major direction is accepted, run second-thought critique automatically.
Major direction includes:
- finalizing a design/plan
- taking a meaningful tool call or action
- committing to an execution path that is costly or hard to reverse
Hybrid Trigger Model
Use a hybrid model:
- Automatic gate at high-commitment moments.
- Manual invocation available at any time.
Automatic gate should trigger when requests involve:
- committing/merging/shipping/deploying/publishing
- meaningful file edits, deletions, migrations, or refactors
- potentially destructive or irreversible actions
- high-cost, high-blast-radius, or under-specified execution steps
If a platform supports runtime message transforms or pre-action hooks, use those to trigger the gate automatically. If not, enforce this via startup instructions and agent process discipline.
Escape Hatch
Allow one-shot bypass when speed is intentionally prioritized:
- User can include
[st:skip <reason>]or#st-skip. - Agent should acknowledge the bypass reason explicitly.
- Bypass should not become default behavior.
Priority
- User instructions
- Second Thought skills
- Default behavior
How to use
- Capture/infer
original_intent. - Summarize/infer
current_context. - Build critique payload automatically.
- Run the critique engine.
- Follow
statusstrictly:
proceed-> continue with checkpoint disciplinerevise-> improve proposal and rerun critiqueprompt_human-> askprompt_to_humanbefore continuing
- Present results in conversational human language by default (few sentences); keep raw JSON internal unless explicitly requested.
Platform mapping
If skill/tool names differ by platform, use the mapped equivalents in references/:
codex-tools.mdcopilot-tools.mdgemini-tools.md
Anti-patterns
- treating agreement as proof
- validating the latest suggestion without checking original intent
- pushing forward with high-impact ambiguity
- giving generic caution without actionable guidance