thinking-fast-and-slowlisted
Install: claude install-skill Trustbaselab/trustbase-skills
# The AI System-1 Trap · A System-1/System-2 Review Protocol
**Kahneman's framework:** **System 1** is fast, automatic, associative, pattern-matching, and aims at *coherence*; **System 2** is slow, effortful, and responsible for review — but by default it is **lazy** and simply adopts System 1's conclusions. Transplant those two into an AI workflow: the default output path of an LLM/agent (next-token associative generation) is **functionally ≈ System 1** — fluent, coherent, confident, and not necessarily correct. All this skill does is **turn System 2 into a process**, not a slogan.
> Function-analogy disclaimer: AI has no biological cognitive resources. "AI's System 1/System 2" is an **analogy used to design a review process**, not a neuroscience claim (see the boundaries table).
## When to use
- Reviewing AI/agent output: the user asks "is this conclusion reliable", "is this a hallucination", "that answer felt too smooth"
- Answers that move with the question: swapping framing (gain/loss), word order, or supplying a "reference value" makes the answer drift
- Making estimates and plans: schedule, cost, success probability, risk probability — especially when only the inside view is available
- Attribution and retrospectives: turning a success or failure into a causal story; "this failed because of X"
- Before an irreversible decision: dropping a database, changing an architecture, a large investment, a release — needs calibrated output and red-team review
- Agent self-che