Menu
Show all learning-loop commands with usage and modifiers. Usage: /learning-loop:help (no args).
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Deep research with the local librarian doing the token-heavy middle. Scope on Claude, Search+Fetch+Extract on the local Ollama model, adversarial Verify + Synthesize back on Claude. Usage: /learning-loop:research "<question>". Falls back to Claude-native WebSearch when the librarian is unavailable or sub-tier.
End-of-session consolidation. Usage: /learning-loop:reflect (no args). Reviews conversation, extracts learnings, routes to auto-memory or vault, cross-links projects, promotes inbox notes. Run after substantial work sessions.
measure whether a /dream pass helps, hurts, or ties against no consolidation. usage: /learning-loop:dream-eval --mode={single|control|repeated} [--passes=N] [--mine]. single scores pre/post on the live corpus; control forks two clones and dreams one; repeated dreams a clone N times and plots the drift + content-survival curve.
Consolidate auto-memory between sessions. Usage: /learning-loop:dream (no args). Four-phase cycle: Orient, Gather Signal, Consolidate, Prune Index. Seven operators: MERGE, RESOLVE, ABSTRACT, COMPRESS, PRUNE, LINK, DATE NORMALIZE. Rebuilds MEMORY.md.
Cross-store correction for a retracted or updated belief. Usage: /learning-loop:rewrite "old pattern" "new pattern" [reason]. Searches vault, auto-memory, and episodic history; presents an impact map for triage; executes user-approved changes; records a supersession so future episodic searches get annotated.
Diagnose your learning-loop installation. Runs health checks, presents issues, offers per-fix remediation, re-runs each check after the fix to confirm. Safe to run anytime; only makes changes you approve. Pass --redact to scan plugin data files for leaked credentials instead.
| name | help |
| description | Show all learning-loop commands with usage and modifiers. Usage: /learning-loop:help (no args). |
Present the guide below when the user runs /learning-loop:help or asks what the learning loop can do. Adapt the level of detail to context: if they seem experienced, lean on the quick reference at the end. If they're new, walk them through the narrative.
/learning-loop:help: show all commandsPresent this guide:
The learning loop turns conversations into lasting knowledge. Ideas start rough, get refined through research, and settle into permanent notes in your vault. Every command serves one of three jobs: bringing ideas in, making them stronger, or keeping things tidy.
Curious about something? Run /learning-loop:discovery.
/learning-loop:discovery "spaced repetition"
It searches your vault for what you already know, researches the web for what you don't, and walks you through the topic interactively. You steer: it digs. At the end, key insights land in your inbox as atomic notes.
Want to just browse without saving anything? Add --surf:
/learning-loop:discovery "spaced repetition" --surf
Other options: --style guided|branch|checkpoint, --tone academic|conversational|persona.
Need depth, not a walk-through? Run /learning-loop:research.
/learning-loop:research "does creatine help cognition in sleep-deprived adults?"
Deep research with the local librarian doing the token-heavy middle: Claude scopes the question into search angles, the local Ollama model (12b+ tier) runs Search, Fetch, and Extract, then Claude adversarially verifies the surviving claims and synthesizes a cited report. Falls back to a Claude-native research path automatically when the librarian is unavailable or the model is below the research tier.
Reading something good? Run /learning-loop:literature.
/learning-loop:literature https://example.com/article
It fetches the source, extracts the core ideas, and writes a literature note to 2-literature/. The source's ideas, captured clean: your commentary goes in separate notes that link back.
Need a quick answer? Run /learning-loop:quick.
/learning-loop:quick "how much have jaguar prices dropped recently?"
It searches your vault and the web in parallel, verifies key claims, and gives you a direct answer in 3-10 sentences. If the answer contains something novel and durable, it auto-captures a note to your inbox. No interactive rounds, no steering -- just a fast, sourced answer.
Quick capture mid-conversation? Run /learning-loop:quick-note.
/learning-loop:quick-note "insight as title"
It grabs the insight, finds vault links, and drops an atomic note in 0-inbox/. No preview, no approval: just a one-line confirmation and back to work. Run it with no args and it infers the insight from conversation context.
Finishing a work session? Run /learning-loop:reflect.
It reviews what happened in the conversation, extracts anything worth keeping, and routes it to the right place: behavioral stuff to auto-memory, knowledge to your vault. This is how the loop closes. Without it, insights from the session evaporate.
Notes land in 0-inbox/ as rough captures. Two commands move them forward:
/learning-loop:deepen takes a single note and strengthens it. It reads the note, scores its maturity, researches what's missing, rewrites it in vault voice, and promotes it when ready. Shallow notes get heavy research; deep notes get a light touch.
/learning-loop:deepen "note name"
/learning-loop:verify assesses both quality and source integrity. It scores each note on depth, sourcing, linking, and voice, then checks that cited sources are real and claims match what they say. Use it to find where to invest /deepen effort and catch fabricated references.
/learning-loop:verify inbox
/learning-loop:verify "distributed systems"
/learning-loop:verify permanent
/learning-loop:gaps is the scientific method applied to your vault. It doesn't just find what's missing: it challenges what you believe. For any topic, it extracts your vault's claims, searches for counterarguments and criticisms, and surfaces tensions, absences, and thin ice. Findings are framed as questions, not verdicts. You decide what they mean.
/learning-loop:gaps "theanine"
/learning-loop:gaps
/learning-loop:gaps --sweep
Focused mode analyses a topic. No-argument mode auto-picks your densest unchallenged knowledge cluster. Sweep mode runs across the entire vault. Depth scales to note maturity: permanent notes get deep scrutiny. Counterpoint notes land in your inbox like any other knowledge, tagged #counterpoint and linked back to the challenged note.
/learning-loop:rewrite is cross-store correction. When a belief turns out to be wrong, the old version sits in three places: vault notes, auto-memory, and episodic history. This skill edits all three coherently. Vault and auto-memory are rewritten or archived; episodic history is annotated via a supersession record so future retrievals carry the correction inline.
/learning-loop:rewrite "old pattern" "new pattern"
/learning-loop:rewrite "old pattern" "new pattern" "explicit reason"
/learning-loop:rewrite
No args: infers the change from recent conversation context. With args: searches for every note, preference, and past conversation that encodes the old pattern, presents an impact map for triage, then executes the approved changes and records a supersession.
/learning-loop:inbox is batch triage. It reads every note in your inbox, clusters them by topic, and recommends actions: promote, merge, deepen, or delete. Promotions happen automatically. Merges and deletes wait for your approval.
/learning-loop:health is your vault's status check. It scans for ghost duplicates, near-duplicate pairs, orphan notes, stale inbox entries, embedding gaps, and broken links. Light mode (default) gives you counts in seconds. --deep launches full analysis. --auto fixes the safe stuff without asking.
/learning-loop:health
/learning-loop:health --deep
/learning-loop:health --deep --auto
If you've enabled the librarian (see below), /health also shows pending librarian observations in the dashboard. Run /health --librarian to review and act on them.
If /health reports installation problems (missing binary, broken shims, stale daemons), run /learning-loop:doctor — it diagnoses your install and walks you through fixes one at a time, re-checking each after you apply it.
The vault librarian runs in the background when ll-watch is active, using a local Gemma model via ollama: gemma3:12b by default (also the research tier), or gemma4:e2b on 16-32GB machines (triage only). It wanders the vault autonomously, finding orphan notes that should be linked, suggesting tags for under-tagged notes, flagging topic-style titles, flagging duplicates against near-neighbours, and marking potentially stale claims. It queues observations; you review them with /health --librarian.
/learning-loop:health --librarian
Phase 1 presents link suggestions, tag suggestions, voice flags, and duplicate flags for approval. Phase 2 hands staleness suspects to Claude for deep investigation. Enable via /init Phase 7 (requires ollama + 16GB+ RAM). No API calls, completely local and free.
/learning-loop:refresh is pure recall: what does your vault already hold on a topic? No research, no new notes. Just surfaces what's there, organized by knowledge depth, and suggests what to do next.
/learning-loop:refresh "authentication patterns"
/learning-loop:ingest pulls bulk context from external systems into your vault and auto-memory. Four modes supported:
/learning-loop:ingest linear # my assigned tickets
/learning-loop:ingest linear "AI Assistant" # tickets from a specific project
/learning-loop:ingest repo ~/dev/acme/monorepo # scan a repo
/learning-loop:ingest repo ~/dev/foo --deep # force the parallel deep fan-out
/learning-loop:ingest context # paste text to extract
/learning-loop:ingest bundle <path> # restore a harvest bundle
It fetches the data, extracts atomic insights, previews them for your confirmation, then routes project-state to auto-memory and durable insights to 0-inbox/. Run it when starting a new project, onboarding to a codebase, or pulling in work context from Linear.
For repo, a Haiku gate decides between a single-pass scan and a 5-wide parallel fan-out (4 deep mappers covering stack/architecture/conventions/domain + 1 state sidecar) that stages structured docs at <vault>/_ingested-repos/<slug>/ and synthesizes them into atomic notes. Pass --deep to skip the gate and force the parallel path.
For bundle, point it at a harvest-bundle-<date>/ directory that /learning-loop:harvest emitted on another instance you own. It restores the carried auto-memory files and vault notes verbatim: notes land in 0-inbox/ for normal triage, existing files are never overwritten, and conflicts are listed for manual merge.
first time → /learning-loop:init → vault path, persona, folder structure
federation → /learning-loop:federation → identity, token redeem, visibility, sync test
external → /learning-loop:ingest → auto-memory + inbox notes
curiosity → /learning-loop:discovery → inbox notes → /learning-loop:deepen → permanent notes
deep dive → /learning-loop:research → verified, cited report
question → /learning-loop:quick → answer + auto-capture if novel
reading → /learning-loop:literature → literature notes
mid-work → /learning-loop:quick-note → inbox note (don't break flow)
sessions → /learning-loop:reflect → inbox + auto-memory
cleanup → /learning-loop:inbox → promote, merge, or deepen
recall → /learning-loop:refresh → see what you know
quality → /learning-loop:verify → score quality + check sources → /learning-loop:deepen
challenge → /learning-loop:gaps → counterpoints + rewrites + /deepen queue
correction → /learning-loop:rewrite → vault + auto-memory + episodic supersession
hygiene → /learning-loop:health → diagnose → route to /inbox, /verify, /deepen
librarian → /learning-loop:health --librarian → approve links + tags, review duplicates and staleness
If you've configured federation via /learning-loop:federation (also reachable from /learning-loop:init), search results automatically include notes from connected peers. Peer results appear with a peer:<name>/ prefix in their path.
node ${CLAUDE_PLUGIN_ROOT}/scripts/vault-search.mjs sync refreshes peer indexes from the hubguide/ tree is not shipped with the plugin)| Command | What it does |
|---|---|
/learning-loop:init | First-time setup: vault path, persona, folder structure |
/learning-loop:federation | Configure federation: identity, token redeem, peers, visibility, sync |
/learning-loop:discovery "topic" | Interactive research journey: explore something new or go deeper |
/learning-loop:research "question" | Deep research: local librarian does the token-heavy gathering, Claude verifies and synthesizes a cited report |
/learning-loop:quick "question" | Fast verified answer: vault + web, auto-captures if novel |
/learning-loop:literature <URL> | Capture an external source as a literature note |
/learning-loop:quick-note [title] [body] | Quick capture to inbox: no args infers from context |
/learning-loop:reflect | End-of-session: extract and persist learnings |
/learning-loop:deepen <note> | Strengthen a single note with research |
/learning-loop:verify [scope] | Score quality + verify sources, find what needs work |
/learning-loop:inbox | Batch triage inbox notes |
/learning-loop:refresh "topic" | Surface what you already know: no research |
/learning-loop:gaps "topic" | Challenge vault knowledge: find tensions, thin ice, and missing perspectives |
/learning-loop:rewrite "old" "new" [reason] | Cross-store correction: vault + auto-memory + episodic supersession |
/learning-loop:ingest [linear|repo|context] [--deep] | Pull external context into vault + auto-memory; --deep forces parallel deep mappers on repo |
/learning-loop:seed [--for-job] [--types a,b] [--out <dir>] | Build a portable starter slice for a fresh instance (ramp up at a new job / second machine) |
/learning-loop:harvest [--all] [--out <dir>] | Collect opt-in portable: true, IP-scrubbed insights from this instance to carry back home |
/learning-loop:health [--deep] [--auto] | Vault hygiene dashboard: ghost dupes, orphans, stale notes, broken links |
/learning-loop:health --librarian | Review librarian queue: approve link/tag suggestions, acknowledge voice flags, resolve duplicate flags, investigate staleness |
/learning-loop:doctor | Diagnose your learning-loop installation: runs health checks, presents issues, offers per-fix remediation, re-runs each check after the fix to confirm |
/learning-loop:dream | Consolidate auto-memory between sessions |
/learning-loop:diagram "concept" | Generate Excalidraw diagram for vault |
/learning-loop:uninstall | Guided removal: marketplace uninstall, dependent MCP cleanup, captured-index purge, with per-step confirmation |
/learning-loop:help | This guide |