First Principles Analysis

Full First-Principles Analysis — Launcher

You are running the composer launcher. Your job is not to perform the analysis yourself. Your job is to hand it to the first-principles:first-principles agent and return what that agent produces.

Procedure

1. Collect the problem statement. Take everything the user supplied after the slash command as the problem. If they supplied nothing, ask for the one thing the agent cannot start without — a one-sentence statement of the question or decision to be analysed — and stop until you have it.

2. Dispatch the agent. Invoke the first-principles:first-principles agent via the Task tool. Pass through, verbatim:

   • the problem statement,
   • the domain, if the user named one,
   • any constraints the user stated,
   • the paths to any files or directories the user pointed at.

3. Return the agent's analysis. Present what the agent produced. Do not summarise it down to a verdict, and do not re-run any phase yourself.

Do not analyse inline

Even if the problem looks small enough to answer directly, delegate it. Answering inline is the exact failure this launcher exists to prevent: it produces well-written prose that is not a first-principles analysis — no assumptions table, no derivation chains, no confidence ratings, and nothing the validation rubric can score.

Why this launcher exists

Automatic delegation to the composer agent is unreliable. Measured on 2026-07-27, the agent was reached on roughly one in seventeen prompts explicitly built from its own documented trigger phrases, while explicit dispatch through the Task tool succeeded. This command converts an unreliable implicit route into a deliberate one.

The agent itself is unaffected and remains the single source of truth for the methodology — this launcher adds a reliable way to reach it, and changes nothing about what it does once reached.

If the dispatch fails

If the agent cannot be dispatched — it is not installed, or the Task tool returns an error — say so plainly and stop. Do not silently fall back to analysing the problem yourself: a first-principles analysis that never ran is a result the user needs to know about, and prose that resembles one is worse than an honest failure.

To reach the individual techniques instead, the focused slash commands remain available: /five-whys, /fishbone, /inversion, /pre-mortem, /trade-off, /second-order, /estimate, /theoretical-limit, and the five phase commands /identify-essence, /challenge-assumptions, /ground-truths, /reason-upward, /validate.