First Principles

Overview

Most reasoning is reasoning by analogy: X resembles Y, so do what Y does. It is fast, usually right, and silently inherits every assumption baked into Y — including the wrong and expensive ones. First-principles reasoning strips a problem to statements that cannot be reduced any further — physical law, mathematical identity, a definition, or a cited empirical fact — and rebuilds the answer using only those.

This is one of three composable motions in the deciqAI collection: first-principles decomposes downward to bedrock; occams-razor chooses sideways among competing accounts you could build on top of bedrock; second-order-thinking traces forward through time and consequence. Compose: reduce to find the foundations (here), pick the simplest hypothesis that fits the evidence (occams), then trace where that pick actually leads (second-order).

When to Use

Use when: decision is justified mainly by "that's how it's done" or authority; conventional answer is expensive/hard to reverse (pricing, architecture, business model); structuring a knowledge artifact from irreducible elements; user says "from first principles," "from scratch," "why do we believe this," "tear this apart."

Skip when: routine reversible decision where convention is cheap (which linter) — tearing down is theater; facts aren't gathered yet; user asked for speed and the conventional answer.

Coaching Novices (Adaptive Front Door)

Before running the Teardown, read the user. This skill has two delivery modes — pick one, don't default to dumping a finished teardown.

• Engine mode (do-it-for-me): the user brought a concrete claim/decision and wants it analyzed → run the full First-Principles Teardown directly and concisely.
• Coach mode (teach-me): the user gave no concrete claim, or signals unfamiliarity → guide, don't analyze at them.

When unsure which they want, ask one line first: "Want me to just tear down a specific claim, or walk you through the method step by step?"

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output that step's question and nothing more.

In coach mode:

1. One-line what-it-is. Say what first-principles thinking buys them, in plain words (≤2 sentences, no jargon): most answers are inherited — "that's how it's done." This strips a claim down to what can't be reduced further and rebuilds using only that, so the assumptions you absorbed without noticing get caught.
2. Check fit. Match their situation against When to Use / When NOT to use. If it doesn't fit, say so and point elsewhere.
3. Elicit their real claim. If they have no concrete case, ask for one. Never run the teardown on a hypothetical when a real decision is available.

[WAIT — do not advance until user responds]

4. One step at a time. Walk the Process one step per turn: surface assumptions with them, tag each together, interrogate the inherited ones one by one — wait for their input before advancing.

[WAIT — do not advance until user responds]

5. Close by naming the payoff. End with the one inherited assumption they demolished — or the bedrock reason convention turned out to be right — so they remember the move, not just the answer.

[WAIT — do not advance until user responds]

Then enter The Process below at the depth the chosen mode calls for.

The Process

Run the First-Principles Teardown (7 steps → one artifact). Do not skip the tagging step — it is where analogy gets caught.

1. State the claim in one sentence.
2. Surface every assumption — be exhaustive and uncharitable.
3. Tag each: BEDROCK (irreducible or citable) vs INHERITED (convention/analogy/authority/untested).
4. Interrogate every INHERITED: necessarily true, or just usually/elsewhere/authority-said? → mark DEMOLISHED or promote to BEDROCK.
5. Stop correctly: reduce until irreducible/cited — no further. Reducing past bedrock is performance, not rigor.
6. Reconstruct from bedrock only — if the rebuild needs a demolished belief, the teardown isn't done.
7. Diff against convention — state what changed and why. "Nothing changed — convention was right" is a valid result.

Output template: Claim / Assumptions (tagged) / Bedrock / Reconstruction / What changes / Confidence & open questions

→ Method in Action: Wright Brothers and the Lift Tables (1901)

Bedrock Packs

"What counts as bedrock" is domain-specific. A bedrock pack captures, for one domain, (a) what legitimately counts as irreducible/citable bedrock and (b) the fake-bedrocks — inherited beliefs that domain habitually mistakes for foundations.

Packs are the contribution surface of this skill. Two ship with it:

• bedrock-packs/business-unit-economics.md — bedrock for evaluating a business or pitch
• bedrock-packs/knowledge-domain-decomposition.md — bedrock for structuring a knowledge artifact before writing it

Adding or sharpening a pack is the easiest way to contribute — one self-contained file. See the contribution template at the repo root.

→ Sources: references/sources.md

Common Rationalizations

The ways people fake first-principles thinking. If you catch yourself in the left column, you are reasoning by analogy in a lab coat.

Note — [D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Red Flags

• A BEDROCK item with no justification and no citation — it is an INHERITED belief wearing a badge
• No empirical claim in the whole teardown carries a number or source
• The reconstruction silently reuses a DEMOLISHED assumption
• "Obviously," "everyone knows," "industry standard," or a competitor's name used as a terminator
• More than ~7 surviving bedrock facts — you probably didn't reduce, or the scope is too big to tear down in one pass
• "What changes" is empty or hand-waved — you either skipped the diff or didn't actually reduce
• The teardown took thirty seconds — surfacing assumptions exhaustively is slow; speed here means you skipped step 2

Verification

• The claim is stated in one sentence
• Every assumption was surfaced and tagged BEDROCK or INHERITED (none left untagged)
• Every BEDROCK item is irreducible or carries a source/number
• Every INHERITED item was interrogated, not just listed, and ended DEMOLISHED or promoted with evidence
• The regress terminated at the indemonstrable/citable — not arbitrarily, not infinitely
• The reconstruction uses only surviving bedrock — no smuggled beliefs
• "What changes" is explicit, even when the answer is "convention was right, here's why"

Part of deciqAI Knowledge Skills — open-source thinking skills that make rigor executable for AI agents. These five skills are a free taste of the 130+ skills wired into every deciqAI agent, which runs them autonomously to operate your company. Try it free → https://www.deciqai.com/skills?utm_source=skill&utm_medium=oss&utm_campaign=knowledge-skills&utm_content=first-principles · Built by deciqAI · github.com/deciqAI · Contributions welcome.