---
name: "logic_and_fallacies"
description: "Stress-tests reasoning before it goes out: names common argument flaws (straw man, ad hominem, false dilemma, etc.), runs a check-yourself-first self-audit, and steelmans the other side before rebutting. Triggers on debates, argument evaluation, persuasive writing, reviewing claims, or any 'is this reasoning sound?' question."
metadata: { "includeInPrompt": true }
---

# Logic & Fallacies

## Purpose

Bad reasoning is the most common failure mode in agentic work — and the fallacy label is the most abused correction in discourse. This skill does three jobs in order: (1) audit your **own** reasoning before asserting anything, (2) **steelman** the opposing position before attacking it, (3) name flaws in arguments **precisely** — with the exact break point, never as a drive-by label. The fallacy catalog lives in `references/fallacy-field-guide.md`; a full worked run-through is in `references/worked-example.md`.

## Workflow

### 1. Extract the argument
Write out the argument in this shape before judging it:
- **Claim:** what is being asserted?
- **Premises:** what reasons or evidence are offered?
- **Inference:** how do the premises connect to the claim?

If there is no argument — just a bare claim, an insult, or a preference — say so and stop. Not everything is a fallacy; sometimes it's just an unsupported opinion.

### 2. Check yourself first
Before flagging anything in *someone else's* reasoning, audit your own. Run the **self-check**:
1. What am I treating as true that I haven't verified?
2. What evidence do I actually have, versus what feels true?
3. Would I still believe this if the conclusion were inconvenient to me?
4. What's the strongest objection a smart skeptic would raise — and can I answer it?

This is where motivated reasoning lives: reasoning *toward* the conclusion you want. If you can't pass your own audit, don't proceed to judge anyone else's.

### 3. Steelman before you strike
Construct the **strongest possible version** of the position you're evaluating or opposing — the "steel man," the deliberate opposite of the straw man. Ask:
- What premises must be true for this to hold?
- What is the best evidence a fair opponent would bring?
- Under what boundary conditions would this actually be *right*?

If the steel-manned version holds, say so plainly. Updating your view is not losing an argument — it is the entire point of having one. Charity is not surrender: you steelman to test ideas against their strongest form, not to agree with them.

### 4. Scan for flaws — carefully
Use `references/fallacy-field-guide.md` for the recognition patterns. But **do not pattern-match mechanically.** Before naming a fallacy, check whether context, domain knowledge, or a non-obvious valid inference justifies what looks fallacious — e.g., citing a doctor on medicine is evidence, not appeal to authority; a genuine dilemma is not a false one. The label is only earned when you can point to the exact sentence where the reasoning breaks.

### 5. Deliver the verdict
State one of:
- **Sound:** conclusion follows from true premises; show the inference chain.
- **Flawed:** name the fallacy, quote the break point, and provide a **repaired** version of the argument (strongest valid form of what they were trying to say).
- **Undecidable on current evidence:** say what's missing and how to test it.

## Output Contract
Every analysis reports:
1. The extracted argument (claim / premises / inference).
2. The self-check result (what you verified, what you didn't).
3. The steel man (strongest version of the other side).
4. The verdict: sound, flawed (named fallacy + break point + repair), or undecidable.

## Operating Rules
1. **Accuracy first.** Every factual claim in a skill must be verified via web search or a primary source. Never invent numbers, dates, quotes, or events. Unverifiable = cut it or frame as opinion. No confident unverified claims, ever.
2. Never drop a fallacy name without quoting the exact words where the reasoning breaks. A label without a break point is itself a dodge.
3. Steelman first, always. If you skipped step 3, you didn't earn step 4.
4. Citing relevant expertise is evidence, not "appeal to authority." Popularity, tradition, and emotion are not evidence — unless the question is *about* popularity, tradition, or emotion.
5. Correlation is never causation until a mechanism is shown. One example is never "everyone" or "always."
6. Don't use fallacy-spotting to win. Using "that's a straw man" as a shutdown is performance, not truth-seeking — and the skill is wasted there.
7. Keep `SKILL.md` tight; the catalog and worked examples stay in `references/`.