first-principles
Multi-pass first principles analysis for any domain — business, personal decisions, creative projects, career, health, relationships. Grounds the analysis in real evidence first, decomposes problems to fundamental truths, challenges assumptions through 4 universal lenses (Constraints, Resources, Human, Context), and reconstructs solutions from evidence-backed ground truths. Three depths: lite (quick inline sanity-check), standard, and deep (parallel subagents per lens). Use when the user says "first principles", "from first principles", "analyze fundamentals", "challenge assumptions", "decompose problem", "why does this cost so much", "is this the right approach fundamentally", "what are we really solving", "strip away assumptions", "should I really be doing this". Also use when the user is stuck on a problem that seems intractable, when conventional approaches have failed, when questioning whether the whole approach is wrong, or when facing a "we've always done it this way" situation
kirillgreen · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add kirillgreen/skills --skill first-principles技能说明
name: first-principles description: > Multi-pass first principles analysis for any domain — business, personal decisions, creative projects, career, health, relationships. Grounds the analysis in real evidence first, decomposes problems to fundamental truths, challenges assumptions through 4 universal lenses (Constraints, Resources, Human, Context), and reconstructs solutions from evidence-backed ground truths. Three depths: lite (quick inline sanity-check), standard, and deep (parallel subagents per lens). Use when the user says "first principles", "from first principles", "analyze fundamentals", "challenge assumptions", "decompose problem", "why does this cost so much", "is this the right approach fundamentally", "what are we really solving", "strip away assumptions", "should I really be doing this". Also use when the user is stuck on a problem that seems intractable, when conventional approaches have failed, when questioning whether the whole approach is wrong, or when facing a "we've always done it this way" situation — whether in business, life, or creative work. Do NOT use for quick factual questions, time-critical decisions, trivial/low-stakes choices, or domains already optimized through rigorous first-principles work. Do NOT use when the user just wants a simple opinion or recommendation without deep analysis.
First Principles Analysis
Decompose any problem to its irreducible truths, challenge every assumption, rebuild solutions from verified fundamentals.
First principles thinking — tracing to Aristotle's concept of archai ("the first basis from which a thing is known") — is the practice of breaking a problem into fundamental truths that cannot be deduced from anything else, then reasoning upward from those truths to construct original solutions. It's the opposite of reasoning by analogy ("others do X, so we should do X").
This skill runs a structured multi-pass analysis, not a single-pass template. Each pass examines the problem through a different lens, building a progressively deeper understanding.
Modes
| Mode | Invoke | Approach | Cost |
|---|---|---|---|
| Lite | /first-principles lite | Assumptions map + top-3 challenges + one reconstruction, inline | No subagents, no file save |
| Standard | /first-principles | Sequential passes; devil's advocate as subagent | 1 subagent |
| Deep | /first-principles deep | Parallel subagents per lens + contrarian subagent | 5 subagents |
Every mode starts with evidence intake (Pass 0). The depth of the intake scales with the mode, but no mode skips it — an analysis with no evidence base is a well-formatted guess.
Standard Mode — Passes 0-4
Pass 0: Evidence Intake
First principles analysis run on an empty evidence base produces confident fiction: the model challenges assumptions using its own assumptions, and invents the numbers the gap analysis depends on. Musk's battery insight ($80/kWh commodity floor vs $600/kWh market price) started with real commodity prices, not introspection. Before decomposing, gather the facts the analysis will stand on.
1. Prior knowledge. Search whatever prior-knowledge stores exist in your environment — project docs, a knowledge base, earlier research notes. If the problem concerns a project, read its core documentation. Prior research is priors, not gospel — but starting blind wastes work already done.
2. External facts. Identify the 3-5 load-bearing factual claims the analysis will rest on — costs, prices, market sizes, timings, physical limits — and fetch real numbers with sources using the web tools available (WebSearch/WebFetch or equivalent). A claim that can't be verified in reasonable time stays in the analysis but gets tagged [unverified], so it can't silently harden into a High-confidence ground truth.
3. User-held facts. For personal, career, and life decisions, most of the evidence lives with the user. Ask 2-4 targeted questions (AskUserQuestion) before analyzing — "why do I believe this?" is the user's question to answer, and self-answering it produces a well-structured guess. When the user can't be asked (headless or subagent run), list the questions you would have asked and tag every conclusion that depends on the missing answers as Low confidence.
Output: Evidence Base — the facts the analysis builds on, each tagged by source type:
| Tag | Meaning |
|---|---|
[verified] | Fetched this session with a source, or hard math/physics |
[user-stated] | Provided by the user |
[model-knowledge] | A general fact recalled from training data — plausible but not checked |
[unverified] | A specific claim you tried to source this session and couldn't — caps at Low confidence |
Confidence levels in Pass 3 are computed from these tags. Evidence quality propagates upward; lens agreement does not.
Scale the intake to the problem: a strategic business analysis deserves real market numbers; a personal decision deserves real user answers; neither deserves invented figures.
Pass 1: Decomposition
Strip the problem to its essence. Most problems arrive pre-framed by analogy ("We need a better X" assumes X is the right category).
Step 1 — Restate the problem without any solution implied. Ask: "What outcome does the user actually need?" not "How do we improve the current approach?" If the user says "we need a faster database," the real problem might be "users wait too long for results" — which might not need a database at all.
Step 2 — Surface every assumption. List all assumptions the current situation relies on — explicit and implicit:
| Category | What to look for |
|---|---|
| Industry conventions | "It's always been done this way" |
| Technical constraints | Physics-bound vs. policy-bound vs. habit |
| Economic assumptions | Market prices vs. raw material/fundamental costs |
| User assumptions | "Users want/need/won't pay for X" |
| Organizational ha |