Self-Efficacy and Cognitive-Rot Shielding

Purpose

Every other skill in this pack designs human-AI oversight at the organizational level — routing rules, behavioral specs, accuracy guardrails, override audits, apprenticeship pipelines. This skill is the individual layer underneath all of them: what happens to the practitioner's own judgment, confidence, and thinking quality across hundreds of small moments of "should I think this through myself, or just ask." Accepting an AI's first answer without applying independent judgment, repeated across enough of those moments, produces what the source material calls "cognitive rot" — a slow erosion of the practitioner's own agency and self-belief that no single interaction makes visible, but that compounds. This skill is a standing discipline against that erosion, not a one-time setup step.

Anchored in

Two Anthropic practitioners' accounts, supplied by the user from source video transcripts. Noam Segal, on the mechanism: "cognitive rot... basically this phenomenon where you sort of see the initial output of an AI model and you sort of just accept it, you don't apply your judgment to it... and you slowly let your mind, your involvement in your work, your agency, sort of collapse into this rotting state" — and on why it compounds: "every time you, not the AI, solve a problem and get over a barrier, it increases your baseline of self-efficacy, of self-confidence, of self-belief, and every time you offload that... you're lowering that baseline." Anthropic product lead Dianne Penn, on what a genuine thinking partnership requires instead: "Claude doesn't take over all of my thinking for me... I might come up with my own POV first and then work with Claude through that... what you don't want is an AI that just agrees with you — a thinking partner doesn't just agree with you, it should add to you, and you should come away at the end of the day having better ideas."

Method — a discipline for the human, with an explicit ask of the AI

1. Think-First: never open a blank-page AI conversation for a real judgment call. Before asking an AI to analyze, recommend, or solve something that matters, write at least a few sentences of your own hypothesis, instinct, or draft answer first — even a rough one. This single step is what makes everything downstream a genuine partnership rather than an offload: without a prior POV to compare against, there's no way to notice whether the AI changed your mind for a good reason or just filled a vacuum you hadn't tried to fill yourself.
2. Ask for adversarial sparring explicitly — agreement is not partnership. State the instruction directly rather than hoping for it by default: "Here's my hypothesis. Act as an experienced, critical business consultant. Don't compliment the idea or agree with me — find at least three real logical gaps or risks in it first." An AI's default register tends toward being agreeable; a genuine thinking partner has to be asked to push back, and a response that only ever validates the framing already given to it is optimizing for approval, not for a better answer.
3. Keep voice ownership on anything that has to sound like you. Use AI to edit, restructure, tighten, or fact-check language — not to originate the core argument of something that will be read as your own view (a message, a recommendation, a position). Draft the actual argument yourself first (this follows directly from step 1), then bring AI in for craft, not authorship.
4. Notice the specific signal of drift: relief instead of engagement. A useful, low-effort self-check: after reading an AI's answer, ask whether the reaction was "yes, that confirms/extends what I was thinking" (engaged) or "great, now I don't have to think about this" (offloaded). The second reaction, recurring on judgment calls that actually matter, is the practical early sign of the erosion this skill exists to catch — it's worth naming out loud when it happens, not just noticing it privately.
5. When acting as the AI side of this partnership: reciprocate the discipline rather than defaulting to agreement. If a user presents a fully-formed conclusion and asks only for validation, the more useful response is often to ask what alternatives were considered and what would change their mind — not to withhold help, but because a thinking partner that only confirms isn't actually adding anything. This is the direct behavioral counterpart to step 2: the discipline works only if both sides of the conversation hold it, not just the human remembering to ask for it.

What this skill does NOT do

• Doesn't slow down low-stakes, low-judgment tasks — Think-First and adversarial sparring are for decisions and analyses that matter, not for every routine formatting or lookup request; applying this discipline indiscriminately would itself become a form of unproductive friction.
• Doesn't mean rejecting AI-generated ideas that are genuinely better than the human's first draft — the point is arriving at the comparison honestly (own POV first, then a real critique), not defending the human's initial answer for its own sake.
• Doesn't diagnose or treat a mental-health condition — "cognitive rot" here is a practitioner-skill and engagement framing from AI-industry discourse, not a clinical concept.

Refinement notes

Both source attributions are from named, identifiable Anthropic practitioners describing a real, current internal discourse theme, but the exact wording is drawn from a single transcript source supplied by the user rather than independently cross-verified against a second source — treat the underlying mechanism (offloading judgment erodes self-efficacy; explicit sparring instructions counter default AI agreeableness) as well-grounded, and the exact phrasing as attributed but not independently re-confirmed.

Continue from here

• Organizational-level complement — where individual judgment can never be replaced regardless of AI capability: ../expert-agency-and-apprenticeship-protection/SKILL.md
• Where sycophancy is designed against at the product/system level rather than the individual-practitioner level: ../ai-accuracy-guardrails-and-grounding-design/SKILL.md
• Applying the same "state your own POV first" discipline to reading an AI-generated report or dataset critically: ../../../data-strategy-and-literacy/skills/data-bias-and-quality-critical-reading/SKILL.md
• This pack's shared guardrails: ../../CLAUDE.md

References

• ../../references/hitl-partnership-heuristics-research.md — the pack's shared sourcing and grounding-strength notes
• ../../CLAUDE.md — the pack's shared guardrails

---

Reminder: frontmatter has only name and description. Everything else goes into skills_index.json (run scripts/generate_index.py).