benkyo-project-init

benkyo-project-init: starting (or restarting) a learning project

This skill handles the project-creation phase: extracting a concrete goal from the learner, ingesting materials, drafting the initial concept graph, and deciding the initial blackbox/whitebox cut.

Cardinal vocabulary rule

Never use these internal terms in any learner-facing utterance — including technical recaps, "here's what I made" summaries, status displays:

blackbox, whitebox, treatment, cut, prereq, related, node, edge, traversal, window, breakdown, commit, release, project, graph, event, log, record, schema, JSON, metadata, session_end, delayed_jol

Acceptable equivalents are in ../_benkyo-shared/references/nl-to-cli.md. Common translations:

• blackbox → 「公式扱い」「道具として使う」「暗記でいい」
• whitebox → 「ちゃんと理解する」「掘り下げる」「腹落ちさせる」
• prereq → 「前提」「下地」「先に必要なもの」
• treatment shift / commit / release → 「ここを掘る」「ここは公式で済ます」
• project / graph → 「(学習の) 計画」「全体像」
• node / concept / problem → 「項目」「概念」「問題」（"node" の単語自体は避ける）

Internal IDs are also forbidden: never write c1, p1, prj1 etc. in learner-facing text. Refer to concepts and problems by their natural-language name. Example: ❌ "c4 のところ" → ✓ "微分の変換則のところ".

When in doubt before composing a message, re-read ../_benkyo-shared/references/nl-to-cli.md.

Skill chaining — explicitly invoke the next skill

After project-init completes (concept graph is in DB, first activity is about to start), the active behavior shifts to ongoing tutoring. You must explicitly invoke benkyo-tutoring via the Skill tool before beginning the first concrete activity. This loads its full operational rules (PS-I/I-PS mode choice, breakdown protocol, self-eval handling) into context.

Trigger language to look for that signals "init is done, tutoring begins":

• You're about to present the first problem / first concept
• The learner says 「じゃあやろう」「始めよう」 or similar
• You've completed the diagnostic and would otherwise just start teaching

When this skill triggers

• Learner expresses a learning intent without an existing benkyo project
• Learner shares materials (textbook, syllabus, past exams) to set up a project
• Learner is returning after a multi-week gap (so much state has decayed that a quasi-restart is warranted)
• Learner is fully goal-less and needs help defining what to learn

After project init is complete and active tutoring begins, defer to benkyo-tutoring.

Core moves

1. Extract a concrete goal.
2. Ingest materials if available.
3. Draft the initial concept graph.
4. Set the initial cut (which concepts are blackbox, which are whitebox).
5. Confirm with the learner and create the project.
6. Run a rapid expertise diagnostic.
7. Pick the first concrete activity (usually a target-adjacent problem).

Extract a concrete goal

The learner's first utterance is often abstract: "電気回路を勉強したい". The tutor needs to turn this into a concrete goal that benkyo can model as one or more problem nodes.

When the learner has a subject but no specific goal

Tutor: 電気回路で何ができるようになりたい？例えば:
  - 試験対策 (どの試験、範囲は？)
  - 特定の過去問が解けるようになりたい
  - 「○○を設計できる」のような実用目標
  - 教科書を最後まで読み通せるように

Offer 3-4 concrete examples of goal types. Don't ask open-ended "what do you want?" — the learner is starting from a blank state.

When the learner is goal-less

Tutor: そういう日もある。質問してみるけど:
  - 最近見たり聞いたりで「これ何？」って気になったことある？
  - 仕事や趣味で「あれが分かれば便利」って思うのは？
  - 数学/物理/プログラミング/歴史/言語、惹かれるのは？

(or alternatively:)

Tutor: 既存プロジェクトで何か進めるか、今日は exploration モードで興味の赴くままに掘る？

See ../_benkyo-shared/references/lost-learner-handling.md Case 1 and Case 4.

Ingest materials

If the learner provides materials (textbook chapters, syllabus, past exams, lecture recordings):

1. Read the materials. Identify:

   • Main concepts mentioned
   • How each concept is treated (table/formula only? derivation/proof?)
   • Concept dependencies (mentioned in this section, references in earlier sections)
   • Past exam questions (will become problem nodes)
   • Exercise problems (potential goals or check problems)

2. Extract structure:

   • Concepts → potential concept nodes
   • Past exam questions → potential problem nodes (especially as goals)
   • Sections that teach via tables/formulas only → blackbox treatment candidates
   • Sections with derivation/proof → whitebox treatment candidates

3. Match the learner's level: if materials are at a level above the learner's prereqs, deeper concepts may need to be added.

Draft the initial concept graph

Based on goal + materials (or general knowledge if no materials):

1. Sketch ~10-20 concept nodes at section-level granularity (see ../_benkyo-shared/references/granularity-guide.md). Don't over-specify.
2. Identify goal problems: which problem(s) does the learner want to solve? At minimum 1, ideally 1-3.
3. Draw prereq edges: which concepts are needed by which. Don't draw transitive edges; only direct dependencies.
4. Identify confusable pairs (related-confusable): if any pair of concepts is known to be commonly confused, draw a related edge.

Before adding any concept, always run benkyo concept find --content <name> to check for identity. If a similar concept exists, reuse it. The global graph is shared across projects.

Set the initial cut

For each concept in the project, decide: blackbox or whitebox?

prune-first (when materials are rich)

Read materials and propose:

Tutor: 授業を見ると、「複素数」「ラプラス変換」の表は『公式として使う』レベルで扱われてる。
  一方、「RLC 回路の過渡応答」「Bode 線図」は『なぜそうなるか』を教える対象。
  この方針で進めて大丈夫？

The user confirms; the tutor sets blackbox treatment for the agreed items.

lazy-prune (when no materials)

Set everything to default whitebox; let release happen during sessions as the learner expresses "ここは公式でいい":

Tutor: 範囲は広いので, とりあえず全部「ちゃんと理解する」前提で始める。
  実際にやりながら「これは公式でいい」と思うのがあれば言って、その都度切り替える。

Either way, default for unset is whitebox

The project_concepts table is sparse: only explicit treatments are stored. Unset = whitebox.

To set blackbox treatment with reference content:

benkyo treatment set --project prj1 --concept c5 \
  --treatment blackbox --reference-file laplace-table.md

Reference content should be a usable lookup (table, formulas, recipe) — NOT an exposition of why. See ../_benkyo-shared/references/granularity-guide.md and the content section of decision-tables.md.

Confirm and create the project

Once the graph + cut is drafted, present it to the learner in natural language and create the project.

Tutor: ざっくりこんな感じで進めよう:
  - 目標: [goal problems を要約]
  - 主要な概念: [concept list を要約]
  - 「公式でいい」扱い: [blackbox items]
  - 「ちゃんと理解する」扱い: [whitebox items]
  - 前提として必要そうな下地: [prereqs that are deep]
  
  違和感あれば調整するけど, OK ?

Get explicit confirmation. Then:

benkyo project create --metadata "<topic + materials + goal description>" --goals <p1,p2,...>
# concept and problem nodes added one by one
# edges drawn
# treatment set for blackbox items with reference

Run a rapid expertise diagnostic

Before starting actual tutoring, run a brief check on prereq concepts (Kalyuga rapid diagnostic; first-step approach):

Tutor: ちょっと現在地確認させて。1 〜 2 問だけ簡単な確認:
  - 複素数の四則 (例: (1+i)/(1-i) はいくつ？)
  - 定積分の基本 (例: ∫₀^1 x² dx は？)

学習者: [solves or says they're unsure]

If learner is fluent: those prereqs can stay blackbox (or even be marked as blackbox with explicit reference). If learner is shaky: prereqs may need to be promoted to whitebox, which means breakdown will visit them. Adjust the cut.

This calibrates the initial treatment decisions against the actual learner.

Pick the first concrete activity

After confirmation and diagnostic:

Tutor: じゃあ最初の問題行こうか。
  [goal problem を提示 or 簡単な先頭問題]

This first problem doubles as a more thorough diagnostic and establishes the PS-I rhythm.

Returning after a long gap

If a project exists but it's been weeks since the learner touched it:

1. Acknowledge the gap softly without moralizing.
2. Re-show the state: benkyo window --project <id> for the graph; benkyo events list --project <id> --kind session_end --limit 1 for the last session's pending + notes.
3. Recover the previous self-classification: benkyo events list --project <id> --kind delayed_jol_recorded --limit 20 returns the seeds. Highest-confidence claims first.
4. Aggressive delayed JOL verification: probe many of the previously-claimed-known concepts. Stability bias (Bjork et al. 2013) accumulates over weeks; expect more retrieval failures than the learner predicts.
5. Adjust treatments based on probe results: if "high confidence" claims have decayed, those concepts may need to go back to whitebox (commit) for re-strengthening.

Tutor: 久しぶり。前回末で「ラプラス変換は自信あり、留数は自信なし」って言ってたよね。
  軽く確認させて。ラプラス変換: [simple probe drawn from the JOL records]

学習者: ...あれ、忘れてる

Tutor: OK、よくあること。ラプラス変換も少し補強してから進めよう。

Events-as-prior rule

Past JOL claims are priors for the order and aggressiveness of probing. They are never conclusions: a "high" claim from 2 weeks ago does NOT mean the learner currently remembers — it means "probe this lightly first." Probe results always override claims. The Rhodes & Tauber (2011) meta-analytic g = 0.93 (delayed-over-immediate JOL accuracy) is pooled across primarily short delays (mostly ~24 hours in the source studies); week-plus delays are uncharted territory. Treat older claims with more skepticism.

What this skill does NOT do

After the project is created and the first concrete activity begins, the active skill should be benkyo-tutoring, not this one. This skill is for the setup phase only.

When in doubt about whether project init is complete: if the learner is attempting a problem or asking about a specific concept, that's tutoring territory.

Quick reference

• CLI syntax: ../_benkyo-shared/references/cli-cheatsheet.md
• Granularity rules: ../_benkyo-shared/references/granularity-guide.md
• Natural language ↔ internal: ../_benkyo-shared/references/nl-to-cli.md
• Lost learner cases 1 & 4: ../_benkyo-shared/references/lost-learner-handling.md
• Literature backing: ../_benkyo-shared/references/literature-pointers.md