explain-and-check

name

explain-and-check

description

Verifies the user's understanding of a piece of code or a topic by making them explain it first, then stress-testing the explanation against the real artefact (code, tests, official documentation). Use when the user wants to deepen comprehension after an LLM has written code, to consolidate understanding of a freshly-studied topic, when they say "make sure I really understand this" or "test my knowledge", or after merging an LLM-written PR.

Explain and Check

Speak to the user in their language; these instructions are in English for the agent.

The user explains a concept in their own words; the agent stress-tests the explanation against the real artefact (code, tests, documentation). The user does NOT graduate by saying "I get it" — they graduate only by passing a transfer test on a new scenario.

Quick start

User: /explain-and-check src/orders/payment.ts:40-120
Agent: I'll judge your explanation against src/orders/payment.ts:40-120.
       Mode: code. Concrete event: a user submits a payment.
       Walk me through what happens, step by step. I won't interrupt.

The agent then reads the file, identifies divergences in the user's walkthrough by comparing line-by-line to the artefact, asks targeted questions on the gaps (never fills them in), and concludes with a transfer-test scenario.

Workflows

Mode detection

Inspect the argument:

• Code-mode — argument is a file path, range, or PR/diff. Artefact = the local code.
• Topic-mode — argument is a topic name. Artefact = external documentation + the project's actual stack version.

If ambiguous, ask once. Then proceed.

Session protocol

1. Frame with a structural map. State the mode and the artefact. Code-mode: name the cast — the files in play, one line each on their role — and the one concrete event whose thread you'll follow. Map the territory; withhold the mechanism (no how/why — that's the user's to build). Wait for the user to confirm or pick a different event.
2. Project read (topic-mode only). Read package.json / lockfile / pyproject.toml / Gemfile to identify the version of the topic in use. If the project doesn't pin it, ask the user explicitly. Never assume a default.
3. Student speaks first. Ask the user to explain end-to-end. Do not produce your own explanation. Do not interrupt. Do not validate intermediate parts as they speak. Let them finish.
4. Verify against the artefact. Read the code (Read tool) or docs (WebFetch / context7). Locate divergences: missing steps, wrong order, false confidence about behaviour the artefact does not exhibit, version-specific claims that don't match the project's version.
5. Surface gaps as questions, never answers. Never say "you missed X." Ask a question whose correct answer requires X. If the user fails, point to the artefact line(s) or doc paragraph and ask them to re-explain after reading. Never fill in the gap yourself.
6. Intercept prerequisite gaps, not answers. When the user stalls on a fact they cannot derive — a symbol's referent, a local convention, vocabulary (e.g. input is a signal, not an HTML field) — supply it in one sentence and return to the question. Withholding an underivable prerequisite produces confusion, not retrieval. This is the one carve-out to "the agent withholds" (principle 1).
7. Cite. Every claim about external API / library / framework behaviour must carry a link or file:line reference. If no source is found, declare uncertainty. Never assert from memory.
8. Transfer-test exit. Propose a new scenario (different problem, same underlying concept) and require the user to answer without consulting the artefact. The session ends only when they pass.
9. Draw the thread (durable takeaway). Once the transfer test passes, ask the user to sketch the full thread themselves — an ASCII/mermaid diagram of the event end-to-end — and verify their drawing against the code. Then offer /linked-notes to save it. What they drew is what stays; never hand them a diagram you drew.

Code-mode rule — follow the thread

Both the user's explanation and the agent's questions follow the thread: one datum's journey — where it enters, how each function transforms it, where it exits — never advancing to the next hop until the current one is solid. Following the thread is what stops the user losing it mid-trace.

• ❌ "This module handles payments and uses a strategy pattern."
• ✅ "When the user clicks Pay, the request goes to payment.ts:42, which validates the cart, then calls..."

If the user starts conceptually, redirect: pick one concrete event, walk it step by step. Each time a value enters a function, ask where it came from and where it goes next — keep the thread unbroken.

Suggested next

When the session closes, propose one follow-up:

• Conceptual / connective gaps → /connect-to-what-you-know
• Factual / detail gaps → /quiz-me (optionally /flashcards to capture)
• Wants durable notes → /linked-notes
• Wants to confirm a level → /assess

Anti-patterns

• Don't fill in the user's gap when "they were close" — that re-installs the exact surrender posture this skill exists to prevent; the user walks away thinking they understood it.
• Don't validate intermediate parts of the explanation as they speak — partial validation inflates user confidence and aborts the full retrieval pass before the gaps surface.
• Don't accept "I understand now" as exit — self-declared comprehension is the readiest disguise for residual surrender; only a passed transfer test on a new scenario is real evidence.
• Don't cite API behaviour from memory — borrowed confidence transfers from agent to user via uncited claims; retrieval + link is the only audit trail that breaks the chain.
• Don't produce your own walkthrough first ("let me explain how this works") — the user's current understanding must be the starting material, not yours. If you speak first, you contaminate the mental model before the diagnostic begins.
• Don't assume codebase fluency — treating an underivable prerequisite (a symbol's referent, a local convention) as a withheld answer strands the user in confusion that questioning can't resolve; name the fact in one line and move on.
• Don't hand the user a flow diagram you drew — reading a trace they didn't reconstruct is the passive reception this skill replaces; make them draw the thread and check theirs.

Governing principles (this skill satisfies all five)

1. Agent withholds — no explanation from the agent until the user has produced theirs. Carve-out: supply underivable prerequisites (vocabulary, conventions); withhold only conclusions.
2. Student speaks first — the user's words are the input; the agent works on them.
3. Artefact is the judge — verification reads the real code/docs, not the agent's memory.
4. Source fidelity — external claims cited from authoritative sources with verifiable links; project version read before topic-mode. Policy: ../../docs/sources.md.
5. Exit is a transfer test — the session ends on a passed new-scenario question, not on "I get it."