| name | learn-anything |
| description | Learn any topic 10x faster using the Specificity Method โ a 5-step loop that turns "I don't know what I don't know" into a verified, tested mental model. Use when the user is confused, stuck, says "I don't get X", "I don't know what I don't know about X", "help me understand X", "teach me X", "I want to learn/grok X", or expresses vague frustration or confusion about a concept/architecture/tool and wants to truly understand it rather than just get a quick answer โ even if they never use the word "learn". This orchestrator owns the full journey: prefer it over the individual learn-articulate / learn-decompose / learn-verify / learn-reconstruct / learn-apply-test step skills whenever the user is starting fresh or wants end-to-end understanding. It runs those five as a loop. |
Learn Anything โ The Specificity Method
Move the user from a fuzzy "I don't know what I don't know" to a specific, verified, tested mental model. Specificity is the whole game: a vague feeling cannot be solved; a precise "I don't know X" can.
This skill is the orchestrator. It runs five steps as a loop and carries state between them.
The loop
1. ARTICULATE โ turn the fog into one specific "I don't know X"
2. DECOMPOSE โ break X into pieces; mark each known โ / unknown โ
3. VERIFY โ run the 6 accuracy checks on every โ
4. RECONSTRUCT โ user explains it back in their own words; you rule fake vs real
5. APPLY+TEST โ Predict โ Run โ Compare โ Fork
match โ understood. exit the loop. ship it.
mismatch โ a precise gift. re-enter at step 2 (re-decompose) and loop.
Failure is not a fallback โ it is part of the system. A mismatch tells you exactly which piece was misunderstood. Loop back.
How to run it
- Create the state file
learn-state.md in the current working directory (template below). It is the loop's memory.
- Dispatch each step to its sub-skill in order: invoke
learn-articulate, then learn-decompose, then learn-verify, then learn-reconstruct, then learn-apply-test. After each, write its output into learn-state.md.
- At step 5, fork:
- Result matches the prediction โ mark the topic understood, summarize the mental model, stop.
- Result mismatches โ record which piece broke, return to step 2 with that piece, run the loop again.
- Stop conditions: topic understood and a test proves it, OR the user says stop. Never declare "understood" without step 5 passing.
learn-state.md template
# Learning: <topic>
## Gap (Step 1)
I don't know: <one specific sentence>
## Pieces (Step 2)
- [ ] piece A โ unknown โ
- [x] piece B โ known โ
## Verified facts (Step 3)
- <claim> โ sources: <url>, <url> โ confidence: <high/med/low> โ checked: 2026
## My mental model (Step 4)
<user's own words> โ verdict: real / faked-here: <piece>
## Test (Step 5)
- Prediction: <what should happen>
- Ran: <what was run>
- Result: <observed>
- Fork: match โ done | mismatch โ re-decompose <piece>
## Current step: <1-5>
Notes
- Steps 1, 3, 4 lean on existing skills where installed (
superpowers:brainstorm for articulate, deep-research for verify, comprehension-gate/session-teacher for reconstruct). The learn-* sub-skills wrap them with this method's flavor.
- Keep the user in the loop. The point is their mental model, not your prose.
