aspire-to-be-less-wrong — quality + safety report

---
name: aspire-to-be-less-wrong
description: Calibration-first sanity check for any moment when the user is defending a position, auditing a past decision, or asking whether they got something right. Trigger on phrases like "am I sure about this?", "post-mortem", "I was wrong about X", "I'm convinced that...", "self-review", "decision audit", "let me push back", "I'm pretty confident", "we should definitely", "looking back on", "did I call this right?", or any strong-conviction debate where the user might be over-indexed on a prior. Forces the user to (a) state confidence and evidence separately, (b) check whether belief is proportional or inversely proportional to evidence, (c) update toward "less wrong" rather than defend the original call, and (d) distinguish trunk-level mistakes from leaf-level mistakes. Use during retros, design debates, strategy reviews, and any moment of high conviction. Trigger eagerly even when the user does not name Musk or the framework.
---

# Aspire to Be Less Wrong

> "It's OK to be wrong. Just don't be confident and wrong."
> — Elon Musk, *The Book of Elon* (Chapter: Aspire to Be Less Wrong)

## What this skill captures

Physics-style epistemic hygiene. Three load-bearing claims from Musk:

1. **Assume you're wrong.** The default state of any belief is "probably wrong in some way." The job is to be *less* wrong tomorrow than today — not to be right.
2. **Beliefs proportional to evidence, never inversely proportional.** When someone digs in *harder* as evidence against them mounts, that's the failure mode. Conviction must track data, not ego.
3. **Trunk before leaves.** Mistakes at the trunk (fundamental principles) are catastrophic; mistakes at the leaves (details) are cheap. Audit which kind you made.

Being wrong is fine. Being *confidently* wrong is the sin, because confidence is what stops you from updating.

## When to use this skill

- User says "am I sure about this?", "did I get this right?", "I'm pretty confident that..."
- Post-mortem or retro on a decision that went sideways
- Strong-conviction debate where the user is defending rather than examining
- Decision audit before a commit point (hiring, architecture, big spend)
- User just discovered they were wrong and is processing it
- Recurring disagreement where one side keeps doubling down
- Any moment the user uses words like "obviously", "clearly", "definitely", "everyone knows"
- Calibration check after a forecast resolved (right or wrong)

## The how-to

Walk the user through these steps in order. Do not skip. Do not soften.

**1. Force a confidence + evidence split.**
Ask the user to state, separately: (a) their current confidence (0-100%) in the claim, (b) the concrete evidence supporting it, (c) the concrete evidence against it. Most users have never separated these.

> "Aspirationally, you want to believe things proportionate to the evidence. Not inversely proportional to the evidence."
> — *The Book of Elon*, Aspire to Be Less Wrong

**2. Check the proportionality direction.**
Is their confidence going *up* as counter-evidence accumulates? That's the inverse-proportional failure mode. Name it explicitly. Ask: "What evidence, if you saw it, would drop your confidence by 30 points?" If they can't answer, the belief is not falsifiable and the confidence is not real.

> "It's OK to be wrong. Just don't be confident and wrong."
> — *The Book of Elon*, Aspire to Be Less Wrong

**3. Separate trunk mistakes from leaf mistakes.**
Was the wrong call a misunderstanding of a fundamental principle (trunk) or a detail (leaf)? Trunk errors require rebuilding the model. Leaf errors require a patch. Most people pattern-match every mistake to "leaf" because trunk mistakes are humiliating.

> "It is important to view knowledge as a semantic tree. Make sure you understand the fundamental principles (the trunk and big branches) before you get into the leaves (the details)."
> — *The Book of Elon*, Aspire to Be Less Wrong

**4. Ask "less wrong than what?"**
The goal is not to be right. The goal is to be *less wrong* than yesterday's version of you. State explicitly: what did you previously believe, what do you believe now, what is the delta, and what caused the update? If there's no delta, no learning happened.

> "The mental tools of physics are powerful. They tell us to assume we're wrong and that our goal is to be less wrong. Aspire to be less wrong."
> — *The Book of Elon*, Aspire to Be Less Wrong

**5. Prescribe a reading / talking action, not a thinking action.**
Updating in a vacuum is rumination. The bandwidth of reading and of talking to domain experts beats the bandwidth of introspection. Name a specific book, paper, or person the user should consume next.

> "Most people self-limit their ability to learn. It's pretty straightforward — just read books and talk to people."
> — *The Book of Elon*, Aspire to Be Less Wrong

**6. Close with a calibration receipt.**
Restate the user's new confidence (0-100%), the evidence threshold that would move it further, and the date to revisit. Without a receipt, the update will quietly revert to the original belief within a week.

## Common failure modes

- **Performing humility instead of updating.** Saying "I could be wrong" without lowering the confidence number is theater. Force the number to move or do not call it an update.
- **Conflating trunk and leaf.** Treating a fundamental error as a tweak ("I just need to adjust the parameter") guarantees the same mistake at larger scale.
- **Anchoring on the prior.** "I still basically think X, but with caveats." The caveats are the data; the prior should bend, not be wrapped in protective qualifiers.
- **Falsifiability theater.** Naming evidence that would change your mind but choosing evidence you'll never actually observe. The threshold must be plausibly reachable.
- **Symmetric updating against asymmetric evidence.** If evidence is 9-to-1 against, "I'll move 50/50" is cowardice, not balance. Match the update to the data.
- **Reading replaced by debating.** Doubling down in the same conversation instead of going to read the primary source. The bandwidth of speech is hundreds of bits per second; reading is several times that.

## When NOT to use this skill

- Pure execution tasks with no claim being made (e.g., "rename this variable", "format this CSV").
- Crisis moments where speed beats calibration (the building is on fire — do not run a confidence audit).
- Genuinely well-calibrated users on a specific claim — do not force a ritual on someone who already stated confidence, evidence, and falsifier upfront.
- Subjective preference questions (taste, aesthetics) where "evidence" is category error.
- When the user explicitly wants validation for an emotional reason, not epistemic improvement — name that and step aside.

## Source

The Book of Elon by Eric Jorgenson (2026, Scribe Media). Chapter: "Aspire to Be Less Wrong" (in the section "Think Like a Physicist"). Primary source pages 68-72.