Bear hours

Bear Hours

Frame what to learn and why before reading anything. Like YC office hours, but for autodidacts.

Interactive questions

Use the native AskUserQuestion / user-input tool exposed by the current runtime for every framing question.

Hard rules:

• Never list numbered questions in chat — always call AskUserQuestion
• One question per call — a single entry in questions[], never batch Q1+Q2
• STOP after each call; wait for the answer before the next question
• Smart-skip — if the user's initial message already answers a question, skip it
• Title format: Bear Hours — Q{n}/5 : {label}

Open vs fixed questions:

• Open (Q1–Q4): infer 2–3 contextual options from what the user said, plus a final option other — "Other — I'll clarify in my next message". If they pick other, accept a short free-text reply in chat, then continue with the next AskUserQuestion.
• Fixed (Q5): use the explicit options listed below (time box).

Escape hatch: If the user says "skip", "just do it", or "vas-y" — ask at most 2 remaining questions via AskUserQuestion, then write the brief and list gaps under Open questions.

Step 0: Local context scan

Before Q1, ask permission to search the user's machine for prior learning context (inspired by gstack /office-hours). Keep the ask simple — no mention of data, APIs, or providers.

Title: Bear Hours — Local scan

Ask: "Je peux chercher dans tes anciennes études enregistrées liées à ce sujet. OK ?"

Options:

• Yes — search registered studies
• No — start fresh

If No → skip to Step 1.

If Yes → one more AskUserQuestion:

Title: Bear Hours — Scan scope

Ask: "How wide should I search?"

Options:

• A) This repo only — studies/ in the current project
• B) Repo + global index — studies/ + ~/.ours-stack/studies-index.jsonl + ~/.ours-stack/learnings.jsonl (recommended)

Run the scan

Extract 2–4 keywords from the user's topic. Then:

Scope	Actions
A	Glob studies/*/brief.md · Grep keywords in studies/ · read matching brief.md / notes.md if relevant
B	Everything in A · Read ~/.ours-stack/studies-index.jsonl · keep entries matching slug or keywords (title, wedge, repo_name) · skip entries whose study_path is already covered by scope A · read top 2–3 external brief_path matches only · Read or tail ~/.ours-stack/learnings.jsonl (last 30 lines) · keep entries matching keywords

Index-only rule: cross-project discovery reads only ~/.ours-stack/ — never run find on $HOME at runtime. Studies from other repos appear only if registered in the index (via ./setup backfill or bin/ours-stack-register-study after each brief).

Limits: max 3 study folders opened · max 5 index/learnings entries surfaced · framing scan only — do not summarize source material.

Report findings

If anything relevant, show 2–4 bullets in chat. When a past learning applies:

Prior learning applied: [insight] (confidence: high|medium|low, [date])

Use findings to:

• Suggest resuming an existing slug in Step 2 — never auto-resume without user confirmation
• Pre-fill or smart-skip Q3 (current beliefs) when prior briefs already state them
• Opening-message source: if the user already shared a URL, arXiv link, paper title, chapter, repo, or file path in their opening message, record it for direct insertion into ## Source material in Step 3 (no question asked)
• Add ## Prior progress in the brief (Step 3) when scan found context

If nothing found: one line — "No prior local context — starting fresh."

Step 1: Gather context

Ask via AskUserQuestion until all five are answered (or smart-skipped).

Q1 — Topic

Ask: "What exactly do you want to learn? One paper, one chapter, one system, or one concept — not a whole field."

Options: Infer 2–3 concrete topics from the user's message + other.

Q2 — Why now

Ask: "Why now? What decision, project, or deadline depends on understanding this?"

Options: Infer plausible motivations from context + other.

Q3 — Current beliefs

Ask: "What do you already believe about this? What would genuinely surprise you?"

Options: Infer 2–3 beliefs they might hold + other.

Q4 — Entry angle (derives the wedge)

Ask: "Par où tu veux commencer ?"

Options: Three angles instanciés sur le sujet de Q1 (une ligne concrète chacun) + other — "Other — I'll clarify in my next message".

Avant d'afficher les options : si tu n'es pas sûr de nommer une entrée narrative réelle pour l'angle « histoire », fais une WebSearch rapide (1–2 requêtes ciblées sur le sujet de Q1) pour identifier personnage, événement ou anecdote documentée — puis instancie les trois libellés.

Règle « Par une histoire » : jamais de libellé générique (Par une histoire, une anecdote qui incarne le sujet, etc.). L'option doit nommer une entrée narrative réelle et précise liée au sujet — personnage historique, événement, anecdote documentée — dans le libellé même : Par une histoire — <nom précis>.

Modèle (niveau de précision attendu) — sujet « Bitcoin » :

• Les bases fondamentales — comprendre la blockchain, les clés et le minage, brique par brique
• Par une histoire — Satoshi Nakamoto et le bloc genesis (jamais « Par une histoire » seul ni « une anecdote qui incarne le sujet »)
• Vision d'ensemble d'abord — la carte globale du protocole, une verticale à creuser viendra plus tard

Modèle — sujet « électricité » :

• Les bases fondamentales — tension, courant et résistance, loi d'Ohm en premier
• Par une histoire — la guerre des courants, Edison vs Tesla
• Vision d'ensemble d'abord — du générateur au réseau, une verticale viendra plus tard

Adapter chaque libellé au sujet exact de Q1 — pas de formulation générique. Les deux autres angles (bases, vision d'ensemble) restent instanciés comme ci-dessus ; seul l'angle histoire exige le nom nominatif dans le libellé.

After the choice: dériver le wedge concret (sujet + angle) et le confirmer en une phrase dans le chat (ou une micro AskUserQuestion si le wedge reste trop large). Repousser si le wedge couvre tout le champ de Q1. L'angle choisi sera noté dans ## Narrow wedge du brief.

Q5 — Time box

Ask: "How long are you giving yourself?"

Options:

• One session (today)
• One week
• Ongoing (no hard deadline)
• other

Confirm before writing

After Q5, call AskUserQuestion once more:

Title: Bear Hours — Confirm brief

Ask: Summarize topic, wedge (sujet + angle d'entrée), time box, and source material status (verbatim source if captured from the opening message, else TBD — /source-scout) in 3–5 bullets. Include: « Tu as déjà un lien ? Colle-le maintenant, sinon /source-scout s'en chargera. » Then: "Ready to write studies/<slug>/brief.md?"

Options:

• Yes — write the brief
• Adjust the wedge
• Adjust source material

If they pick Adjust source material or paste a URL/title/path in chat at this step, accept it verbatim for ## Source material, then confirm again. If they pick another adjustment option, ask one targeted AskUserQuestion, then confirm again.

Step 2: Choose slug

Derive slug from the topic: lowercase, hyphens, no spaces (per AGENTS.md).

1. Glob studies/<slug>/brief.md (or list studies/ and match).
2. If slug does not exist → use it.
3. If slug exists → read existing brief.md (# title, ## Narrow wedge, Created). Call AskUserQuestion:

Title: Bear Hours — Existing study

Ask: "A study studies/<slug>/ already exists: [title] — wedge: [one-liner]. What do you want to do?"

Options:

• Resume this study — update brief in place
• Same topic, new scope — fork as <slug>-v2
• Different topic — use a new slug (agent proposes 2 alternatives in chat)

4. Never overwrite or resume without explicit user choice (AGENTS.md: ask before reusing an existing slug).
5. Suffix chain: if <slug>-v2 exists, try -v3, then -v4, until a free slug is found.

Step 3: Write brief.md

## Source material rules:

• Default: TBD — run /source-scout
• Opening-message source: if the user's initial message included a URL, arXiv link, paper title, chapter, repo, or file path → write it verbatim instead of TBD
• Confirm-step paste: if the user pasted a link during the confirm recap → write it verbatim instead of TBD

Create studies/<slug>/brief.md with this structure:

# <Title>

**Slug:** <slug>
**Created:** <YYYY-MM-DD>
**Time box:** <duration>

## Why now
<desperate specificity>

## Current beliefs
<status quo>

## Prior progress
<only if Step 0 found context — related slug, brief one-liner, open questions from notes; omit section if none>

## Narrow wedge
<smallest learnable unit — be concrete>
**Entry angle:** <bases fondamentales | par une histoire | vision d'ensemble — one line how this angle applies>

## Success criteria
- [ ] I can explain <X> without jargon
- [ ] I can answer <Y> from memory

## Source material
TBD — run /source-scout

## Open questions
- ...

(Default ## Source material to TBD — run /source-scout; replace with the verbatim source when captured from the opening message or confirm step.)

After writing brief.md, register the study in the global index:

bin/ours-stack-register-study studies/<slug>/brief.md

(Run from the repo root; use the equivalent repo-local path if invoked from elsewhere.)

Step 4: Handoff — passation et enchaînement

After brief.md is written and registered, deliver the closing in three beats. Tone inspired by office-hours Phase 6 Handoff — personalized, direct, no ceremony.

Anti-slop (mandatory):

• Quote the user's actual words for why now (Q2) and entry angle (Q4) — not paraphrased generics
• No « Félicitations ! », no praise paragraphs, no filler
• GOOD: « Tu veux comprendre ça parce que [their why now]. Tu commences par [their angle choice]. »
• BAD: « Votre parcours d'apprentissage est maintenant bien structuré. »

Interdits inchangés : ne pas lire ni résumer de source ici ; ne pas créer proof.md ; rappeler deep over wide seulement si le wedge sonne trop large (une ligne max).

Beat 1 — Ce qui vient d'être créé

One short paragraph in plain language:

Ton étude « <titre> » est lancée. Ta feuille de route est dans studies/<slug>/brief.md : ton sujet, ton angle d'entrée, tes critères de réussite. C'est le document que tous les autres skills liront.

Personalize with the user's exact phrases — their why now and their chosen angle. Never ship the template sentence alone.

Beat 2 — Carte du parcours

Show where they are and what follows (short block, no essay):

✅ /bear-hours — cadrage (fait)
→ /source-scout — trouve 3 sources adaptées à ton angle
  /dense-read — lecture guidée de la source, tranche par tranche
  /study-status — à tout moment, pour voir où tu en es

Beat 3 — Enchaînement direct

Next skill: if ## Source material is still TBD — run /source-scout → /source-scout ; otherwise → /dense-read (source already in the brief).

Call AskUserQuestion once:

Title: Bear Hours — On enchaîne ?

Ask: One line — e.g. « On enchaîne sur <slug> ? »

Options:

• Oui — lance /source-scout maintenant (or Oui — lance /dense-read maintenant when a source is already listed)
• Plus tard — je m'arrête ici

If Oui: immediately Read and follow source-scout/SKILL.md or dense-read/SKILL.md for this <slug> in the same session (equivalent to invoking via the Skill tool). Do not stop after the handoff.

If Plus tard: one line only — e.g. Quand tu veux : /source-scout <slug> or /dense-read <slug>.

Rules

• Do not start reading or summarizing source material here — framing only
• Do not open or inspect source documents even if the user supplies URLs or file paths — record them in the brief only
• Push back on "learn everything about X" — force a wedge via Q4 entry angle
• Never create proof.md (or any standalone proof draft) unless the user explicitly runs /proof-draft or asks to draft public proof
• Cross-project scan: ~/.ours-stack/studies-index.jsonl only — never find $HOME at runtime