chinmay-sawant The rule: if you cannot explain something in a simple way, you do not fully understand it yet. This skill runs that rule as a loop until the explanation holds.
Unslop first
Before any response, run the /unslop skill on your output and apply its fixes. Simple words are the whole point here; an explanation padded with "leverages" and "robust frameworks" fails its own test. Unslop, then send.
Invocation
Parse the topic from the invocation. Three forms:
/feynman <topic>: you explain the topic./feynman this: explain whatever the conversation was just about./feynman me <topic>: role flip. The user explains, you listen and grade.
No topic? Ask once, wait for the answer.
Grounding
Never explain from memory alone.
- Topic touches this repo: read the actual code first. Every claim about it cites
file:line. - Topic outside the repo: read authoritative docs or source before explaining, and say what you read.
- If sources contradict what you believed, trust the sources and say what changed.
The loop
One pass has three moves. Repeat until the exit gate passes.
- Explain it to a smart 12-year-old who shares none of your vocabulary. Short sentences. Concrete analogies. Every technical term either avoided or defined in the same breath it first appears.
- Audit your own explanation for gaps. Label each finding:
- Jargon: a term the 12-year-old would not know, used without definition.
- Hand-wave: "it just handles that", "under the hood". No mechanism named.
- Circularity: X explained by Y, Y explained by X.
- Name-dropping: leaning on a famous name instead of saying what it does.
- Fill each gap at the source: read the code or doc covering exactly that fuzzy spot. Then re-explain, tighter than last time.
Exit gate (all must hold):
- Zero audit findings on the latest pass. Not fewer than last time. Zero.
- Every step names a mechanism, not a reputation.
- The user could retell the explanation correctly without you in the room.
Report the pass count. "Clean in 3 passes" means something; "clean" alone does not.
Grading mode (/feynman me <topic>)
You become the listener.
- Let the user finish a full attempt before interrupting.
- Flag findings with the same four labels above, quoting their exact words back.
- After flags, drill one question at a time: "why does that work?" until they hit bedrock or hit a wall. A wall is the method working, never mock it.
- Point every gap to the file or doc that closes it.
- End by applying the exit gate to their final version and naming what still would not survive retelling.