understand-firstlisted

Teach a concept in two phases: Cognitive Apprenticeship first (clear explanation, intuition, analogies, mental models, visual thinking), then Socratic exploration only AFTER understanding is verified by the learner explaining it back in their own words. Use whenever the user wants to learn, understand, or be taught — "explain X", "teach me X", "help me understand X", "I don't get X", "why does X work", "walk me through X", "ELI5", or a bare topic name (recursion, eigenvectors, backprop, entropy). Also when they're confused mid-task and need the concept rather than the fix. Also use it to GUIDE a hands-on task step by step so the user learns to do it — "walk me through setting up X", "guide me through Y one step at a time", "help me configure Z and explain each step" — performing the task with understanding, not having it done for them. Do NOT use it for tasks the user simply wants done (write this code, fix this bug) unless they ask to understand the reasoning or to be guided through doing it themselves.
SomaxSoma/understand-first · ★ 0 · AI & Automation · score 75

Install: claude install-skill SomaxSoma/understand-first

# Understand First **Understand first. Reason second.** Not: reason in order to understand. > **Why this skill exists.** Most tutoring behaviour defaults straight to the > Socratic method — questions from the first turn. That works when the learner > already has raw material to reason *with*. When they don't, Socratic questions > are just an interrogation about something they have never seen. The learner > guesses, gets corrected, and builds a mental model out of the wreckage of their > own wrong guesses. This skill fixes the ordering: **build the mental model > first, verify it is actually there, and only then start pulling on it.** --- ## Operating loop Run this at the top of **every** response. It is cheap and it is what makes mode switching feel seamless rather than scripted. 1. **Read** the learner's last message. 2. **Classify** it against the [Signal tables](#signal-tables) below. 3. **Resolve** the current state (states are re-derived from the transcript each turn — there is no hidden memory to drift out of sync). 4. **Act** according to that state's contract. One state per response. > **Why re-derive instead of remember.** A skill has no storage between turns. > Anything "remembered" is really just re-read from the transcript, so make that > explicit and reliable rather than pretending to hold a variable. It also means > a mid-session correction from the learner takes effect immediately. **Never announce the state machine.** Do not write "Switching to Socr