learn-skill

learn-skill

Turn "I want to learn X" into a defended, hands-on learning track, so Han never starts from scratch. The skill does the three things Han asked for: research the state of the art, compare the best tools/methodologies and pick for HIS purpose, and lay out a path he can start today and actually understand.

Grounded in research/2026-05-30-learn-skill-methodology.md (the evidence base) and Han's principle: Knowledge x AI = output. AI accelerates the learning; it does not replace the cognitive work. Push back if Han tries to let the AI do the thinking.

The methodology: the Skill-Loop

Five stages, repeated per sub-skill until the project ships. Full rationale + citations in references/methodology.md.

Anchor → Map → Drill → Recall → Explain
(project) (sub-skills (deliberate (spaced  (teach it
          + tools)   practice)  retrieval) back)

Workflow

Run these in order. This is a process skill: do the steps, do not skip the research or the tool comparison.

1. Scope it (do not silently guess)

Establish, by asking Han when unclear (per his "don't silently pick defaults" rule):

• Goal / concrete outcome: what artifact exists when he can do this? (project-first anchor)
• Current level: never-touched, dabbled, or rusty-intermediate.
• Constraints: platform, time budget, hardware, anything off-limits. Pick the obvious default only when options are equivalent; otherwise ask one batched round.

2. Research the state of the art

WebSearch the current (date-stamped) best tools, methodologies, and canonical resources for X. Get real sources. Note what changed recently; tool landscapes move. Do not rely on training-data memory for "the best tool"; verify it is alive in 2026.

3. Compare tools, pick for HIS purpose

Apply the six-criterion rubric in references/tool-rubric.md. Produce a scored comparison table per tool category, then a defended pick tied to the learner's goal (not a generic "best"). Name the runner-up and when it would win instead.

4. Design the path with the Skill-Loop

Using references/methodology.md: anchor on the project, map the sub-skills in dependency order, design deliberate-practice labs (edge-of-ability, tight feedback), mark the durable facts for spaced retrieval, and add a teach-back checkpoint. Order techniques so each builds on the last; no tool dumps.

5. Emit the track

Follow references/track-emission.md exactly to create learning/<topic>/ (README + frontmatter, CLAUDE.md tutoring contract, docs/decisions/ADR-001-framework-choice.md, the structured path, at least one runnable first lab), add the learning/INDEX.md row, and append the _meta/LAB_LOG.md line. Match the repo's existing track voice (read an existing track first).

6. Wire the retention loop

Point Han at the sibling skills that keep it going: anki-builder for the spaced-retrieval deck, concept-explain for ad-hoc concept questions, learning-day-process if he later drops class material, library-cite for book-grounded answers.

Hard rules

• Topic-agnostic. Never hardcode a specific subject into the skill. The same steps serve Rust, RE, statistics, anything.
• Defended picks, not listicles. Every tool in the path has a one-line reason tied to the goal.
• A path he can start today. The first lab must be runnable on day one (install + first win), not week three.
• AI as tutor, not author. The track makes Han do the recall and the building; the AI diagnoses and explains.
• Privacy. Tracks are private (learning/ rules). Only distill to tieubao/til through the privacy gate.

References

• references/methodology.md, the Skill-Loop, evidence, and why-not-the-alternatives.
• references/tool-rubric.md, the six-criterion scoring rubric + worked example.
• references/track-emission.md, exact files + frontmatter to emit under learning/<topic>/.
• sync-to-global.sh, promote this skill from ops-toolkit to the global claude-skills repo so it fires from any repo.