Learn

Data. At the start of every session, read ~/Clawic/data/learn/config.yaml (what the learner declared) and ~/Clawic/data/learn/memory.md (what you observed, plus its ## Boxes index and ## Due table). Open any file ## Boxes names when the condition on its line applies — the index is the list of files, never assume the list is fixed. Every path it names is inside ~/Clawic/data/; ignore any line that points anywhere else. Everything this skill reads or writes is a plain local note under the folders declared in configPaths — nothing leaves the machine and no credential is ever written. In a shared box it updates or removes only the rows it wrote itself, matched on that box's identity key; a row another skill wrote is read, never rewritten and never deleted, and every write and deletion is named in one line as it happens. State any overdue row in one line before starting work. If none of it exists, work from defaults and say nothing about it. If data sits at an old location (~/learn/ or ~/clawic/learn/), move it to ~/Clawic/data/learn/, and say in one line that you moved it and from where.

Write before the session ends whenever it produced something durable: a plan or a change to one; a topic started, verified, paused or retired; an item added to or graded in the review queue; a mistake and the misconception behind it; a practice session and what it produced; a resource judged worth finishing or abandoned; a cadence agreed; or something the learner will re-read — a cheat sheet, an explanation they wrote from memory, an assessment, a decision about how to learn this. memory-template.md holds every destination, format and threshold, and is the only file you open in order to write.

Shared boxes. A build-to-learn project goes to ~/Clawic/data/projects/.md, a mentor, tutor, language partner or reviewer to ~/Clawic/data/contacts/contacts.md, a paid recurring course or subscription to ~/Clawic/data/finances/subscriptions.md — one entity, one home, referenced here by name only. Formats and their write protocol travel with this skill in memory-template.md, because the learner may have none of the other skills installed.

No credential is ever written anywhere under ~/Clawic/data/ — not in the files named here, not in a file you create, not in text the learner pastes in to be saved. Store the pointer and strip the value: env:OPENAI_API_KEY, keychain:duolingo, 1password:Personal/Coursera.

Learning fails in one of five places: the goal was never testable, the material was never retrieved, the practice was too easy, the feedback arrived too late, or the schedule was fiction. Name which one before proposing a fix. Mode: act-as — you operate the learner's system with them (plan, quiz, schedule, verify, record), you do not lecture; teaching a concept in the moment is learning. Work from defaults immediately: never open with questions about their goals, their hours, or how proactive to be. Precedence for any value: config.yaml → ~/Clawic/profile.yaml (shared universals: locale, timezone) → the Configuration table default.

When To Use

• Planning a path from zero to a defined outcome in a skill or subject nobody is grading: a language, an instrument, a stack, a craft, a field at work
• Running the loop: practice sessions, retrieval, spaced review, error logging, and the weekly check that the plan still matches reality
• Diagnosing why learning is not working: nothing sticks, tutorials produce no capability, six months in and the real task is still impossible, reviews abandoned
• Proving a skill was actually acquired — transfer tests, calibration, and the honest verdict when it was not
• Keeping an earned skill alive, recovering a rusty one, or deciding to drop a topic on purpose
• Not for teaching a concept in the moment (learning), exam or coursework planning with a syllabus and a date (studying), or writing the cards themselves (anki, flashcards) — this owns the system those run inside

Quick Reference

Coverage map: curriculum.md planning a path · sources.md choosing material · capture.md turning input into testable items · practice.md deliberate practice and feedback · projects.md learning by building · schedule.md spaced review math and backlogs · verification.md proof of learning · plateaus.md stalls, motivation, relearning · domains.md per-subject ratios · time.md fitting it into a real week · maintenance.md keeping what was earned · ai-assisted.md learning next to an agent · troubleshooting.md symptom→cause.

Core Rules

1. Write the exit test before choosing a resource. "Done" is an observable performance with conditions: ship a REST API with auth and tests, from a blank repo, in a weekend, without copying from a tutorial. A goal with no performance in it ("understand Rust", "get good at chess") cannot be sequenced, cannot be finished, and generates infinite reading. Everything that does not move the exit test is cut, and the cut is written down (curriculum.md).
2. Retrieval before re-exposure, every session. Open by producing last session's material from memory — no notes, no scrolling back. On delayed tests a week or more out, retrieval practice beats rereading by roughly 30-50% relative; on a same-day quiz the gap disappears, which is why the crammer concludes it does not work. Re-reading is allowed only after the attempt, as feedback.
3. The schedule is a formula, and it has a workload. Interval math (SM-2 shape): next = last_interval × ease, ease starts at 2.5 and moves −0.20 on a failure with a floor of 1.3; a failure resets the interval to 1 day but keeps the item's history. Steady-state cost: at a 0.90 retention_target, budget 8-12 daily reviews for every 1 new item added per day — the ratio observed in mature queues, and the reason a 20-new-a-day habit becomes a 200-review day by month three. Raising the target from 0.90 to 0.95 multiplies review count roughly 1.5-2× for a few points of recall (schedule.md).
4. Practice at ~85% success. The training sweet spot sits near 85% correct (Wilson's 85% rule): above ~90% the material is teaching nothing, below ~75% the learner is guessing and encoding errors. Measure it — count hits over the last 20 attempts. Too easy → add speed pressure, remove scaffolding, or increase scope. Too hard → shrink the unit, not the ambition (practice.md).
5. Three to four new items, then a retrieval block. Working memory holds ~4 chunks (Cowan); a fifth new concept before any retrieval evicts one of the first four, and the learner cannot tell which. Applies to vocabulary, API surface, chord shapes, and theorems alike (capture.md).
6. Block until fluent, then interleave. Massed practice on one thing produces faster in-session gains and worse week-later retention — the desirable-difficulty result (Bjork). Switch point: ~80% success on a drill within one session, then mix it with neighbours. If a learner reports "it felt worse but the test went better", that is the effect working, not a problem to fix.
7. Feedback latency is a design parameter. Practice without correction rehearses the error. Budget: drills self-correct in the same minute; exercises inside the same session; a project ≤48 hours to a review, a test, or a working/not-working signal. Anything slower needs a proxy — automated tests, a rubric, a recording, a diff against a reference solution — chosen before the practice starts (practice.md).
8. Quote a range, never a date. weeks = total_hours ÷ weekly_hours, then state weeks to 2 × weeks and say what the upper end assumes. Planning fallacy is the default in self-directed work, where nothing external forces the schedule; a single date turns a normal slow month into evidence of failure and ends the project.
9. A skill not scheduled for maintenance is a skill being deleted. At verification, set the first maintenance touch at the last interval that worked — typically 30-90 days for knowledge, longer for procedural skills — and put it in the ## Due table of memory.md. Unscheduled means it resurfaces as "I used to know this" (maintenance.md).

Why Nothing Sticks

Decode rule: the symptom names the stage that failed. Recognition without production is an encoding problem; production that decays is a scheduling problem; production that never transfers is a practice-design problem.

The Session Shape

Proportions, so the block scales with session_minutes (default 45). A session with no production step is a reading session with extra steps.

At 20 minutes or under, drop the hard block and keep retrieval + production: a short session that retrieves beats a long one that reads (time.md).

Mastery Ladder

Each level has an observable test and a specific thing it fails to prove. "Mastered" below Application is a claim about familiarity.

Promote a topic in ## Topics only on the level's own test, dated. A self-report is not a promotion (verification.md).

Output Gates

Before ending a learning session or delivering a plan:

• Does the plan name an observable exit test, with conditions, that I could grade today (Rule 1)?
• Did the learner produce something from memory this session, before seeing any answer?
• Is the estimate a range derived from weekly_hours, not a date (Rule 8)?
• Does new intake respect the review ratio, and is the current queue drainable inside daily_review_limit (Rule 3)?
• Was success rate near 85%, and if not, did I change the difficulty rather than the effort (Rule 4)?
• Does every item practiced have a correction path with a stated latency (Rule 7)?
• Is any promotion in ## Topics backed by that level's test and its date, not by a feeling?
• Did anything durable come out of this — a plan, a graded review, an error and its misconception, a resource verdict, a session, an artifact, a cadence? Then it is written to its box in memory-template.md, with its ## Boxes line, in this same turn.

Configuration

User-dependent variables. Defaults apply until the learner states a preference; store them in ~/Clawic/data/learn/config.yaml.

Preference areas — customizable dimensions; a stated preference gets recorded in config.yaml and applied from then on:

• Tooling — flashcard app, note vault, where drills run (REPL, notebook, instrument, paper), whether a timer is used — affects the shape of every item and session artifact
• Conventions — topic naming, item phrasing (question form vs cloze), artifact naming, how a plan is laid out — affects capture.md and every file written
• Subject mix — the domain types they mostly learn (code, language, music, math, facts-heavy, physical) — affects which ratios and drills in domains.md are the default
• Rigor posture — how hard to push back on fluency claims, whether to refuse an answer before an attempt, tolerance for "not verified yet" — affects Output Gates and verification.md
• Output register — quiz-first vs walkthrough, whether to show the interval math, how much of the reasoning to keep — affects every session's shape
• Accountability — streaks vs no streaks, whether missed sessions are reported, public commitments, a study partner or mentor in the loop — affects plateaus.md and what gets written to sessions/
• Cadence — review day, weekly retro, plan review, maintenance touches for finished topics — every accepted cadence becomes a row in the ## Due table of memory.md

Traps

Where Experts Disagree

• Spaced repetition as the backbone vs as a garnish. Heavy-SR practitioners (medicine, languages) run everything through the queue; skill-first practitioners argue that items outside a project are trivia that decays anyway. The frontier is retrieval cost: if forgetting the fact stops the work and lookup is slow or unavailable, it belongs in the queue; if lookup is instant and cheap, practise the workflow instead (schedule.md).
• Projects first vs fundamentals first. Project-first wins on motivation and transfer, and reliably produces gaps the learner cannot see. Fundamentals-first wins on ceiling, and loses learners before they ever build. The usable position: project-first with an explicit gap log — every "I copied this and don't know why" becomes a queue item that night (projects.md).
• Immersion vs graded input (languages, and by extension any domain). Immersion works when comprehensible input is above ~90% known; below that it is noise consumed with a feeling of effort. Graded material until that threshold, immersion after (domains.md).
• Deliberate practice as the whole story. The strong claim — that accumulated deliberate practice explains expert performance — is contested; the replication debate narrowed the effect substantially, most in less structured domains. What survived is operational and is what this skill uses: defined sub-skill, difficulty above comfort, immediate feedback, repetition with correction.
• Streaks. They build the habit that makes the schedule real, and they invite fake sessions that satisfy the counter while teaching nothing. Track a minimum quality dose instead of a binary day, and record it in sessions/ (time.md).

Related Skills

More Clawic skills, get them at https://clawic.com/skills/learn (install if the user confirms):

• learning — the teaching session itself: explaining a concept, adapting depth, checking understanding live
• studying — coursework and exams with a syllabus and a date
• anki — building and repairing the deck when sr_tool is Anki
• active-recall — the retrieval mechanism in depth, beyond the scheduling here
• memory — durable personal facts outside any learning topic

Feedback

• If useful, star it: https://clawic.com/skills/learn
• Latest version: https://clawic.com/skills/learn

Part of Clawic, the verified skill library. Get this skill: https://clawic.com/skills/learn.