Cornelius Skills
All 60 skills, grouped by job. Invoke any of them in Claude Code as /skill-name.
Each skill lives in .claude/skills/<name>/SKILL.md in the Cornelius repo. Jump to: Capture · Thinking and deciding · Search and connections · Autonomous thinking · Research · Reference data · Writing and visuals · Brain Dependency Graph · System and maintenance · Shared contracts
Capture (7)
Turn what you read, say, and think into notes — your thinking and external content kept apart.
Extract unique insights and perspectives from your own content (conversations, transcripts, notes). Spawns the insight-extractor sub-agent.
Extract insights from external documents — research papers, books, articles. Spawns the document-insight-extractor sub-agent; requires a session name.
Fully autonomous end-to-end ingestion of a single source (book, paper, article, transcript): prepare to markdown, extract against the live index, refresh the index, auto-link to existing knowledge, changelog.
KB-grounded Socratic interview — searches existing notes on a topic, then runs a one-question-at-a-time dialogue to surface and sharpen your thinking, ending with extract-insights on the transcript.
Review and graduate notes to permanent status using Zettelkasten principles, consolidating AI extractions and document insights into curated permanent notes.
Extract every chapter of an EPUB into separate markdown files.
Extract chapters from PDF, MOBI, and AZW3 books into markdown — TOC-based PDF splitting with OCR fallback for scanned PDFs; MOBI/AZW3 via calibre.
Thinking and deciding (5)
Reason with your own knowledge base.
Solve problems using knowledge-base insights — extracts search terms, runs parallel KB queries, and synthesizes advice grounded in your own frameworks.
Structure a decision, not just advise on it: expand the real option set, classify the decision type, apply the matching decision rule (ergodic filter, expected value, robustness / minimax regret, value of information), and recommend with tripwires.
Two sub-agents believe fully committed positions while the orchestrator performs structural contradiction analysis and synthesis — for stress-testing ideas and high-stakes tradeoffs.
Free-form thinking: consider a topic through successive distinct KB perspectives (hub lenses, original frameworks, wildcard distant notes). One pass per run; state persists; loop-compatible.
Conversational partner mode — the thinking-partner sub-agent embodies the knowledge base as its own memory and engages as an intellectual equal.
Search and connections (6)
Find what you know, and what it connects to.
Fast retrieval — a single FAISS search plus graph connections, no sub-agents or LLM orchestration.
Quick search across the vault using keywords or semantic similarity.
Retrieve relevant knowledge with 3-layer semantic search based on the conversation context.
Discover hidden connections and relationships between notes.
Discover non-obvious cross-domain connections through random sampling and pattern analysis.
Find notes created in the last 14 days and discover their connections to the knowledge base.
Autonomous thinking (5)
The loops that keep working on your open questions — scheduled when Cornelius runs on Trinity.
Autonomous iterative thinking loop — works active topics with rotating analytical moves (ACH, Bayesian updating, steelmanning, cross-domain bridging, implication checks) and persists reasoning state across scheduled runs.
Manage the incubation topic lifecycle — seed new questions, review status, crystallize converged conclusions, retire stale topics.
Autonomous perception layer — scans the KB for new notes matching domain-watch configs, checks gap resonance, probes external signals, and activates strong signals for the incubation loop.
Manage domain-watch configs — seed domains, review scans and proposals, activate proposals into thinking topics, pause domains, update watch queries.
Autonomous crystallization — synthesizes converged topics into AI-inferred notes in a dedicated folder. Never touches the human-curated permanent knowledge base and never changes a topic's status.
Research (3)
Expand the knowledge base from outside sources.
Autonomous research pipeline — discover, extract, and integrate cutting-edge insights into the knowledge base.
Continuous learning heartbeat — autonomously researches, extracts insights, and expands the knowledge base.
Extract the transcript of a YouTube video by URL or ID, with automatic language fallback.
Reference data (8)
Entity records — people, organizations, products, engagements, competitors — in mountable reference scopes. Company is the default scope. Records never become endorsed insights.
Upsert and maintain entity notes in a reference scope, then reindex. Never crystallizes, lifecycle-classifies, or trains on reference data.
The single entry point for incoming information about an entity — resolves it to the one canonical note (never duplicating), reconciles new vs old, upserts, and reindexes.
Structured, temporal lookup (“what do we know about Acme as of today?”) that respects status and validity windows and always prints the as_of date. Read-only.
Read-only integrity report for a reference scope — provenance and enum violations, as_of and supersession integrity, dangling links, orphans, duplicate candidates.
Keep a scope canonical — detects issues via ref-audit, then human-gated fixes: conflicts, bidirectional link repair, and merging duplicate entities.
Staleness sweep against per-type freshness SLAs; market items are re-verified autonomously, CRM items are report-only.
Handle a validity transition (renewal, role change) — creates the new active note, marks the old one superseded, keeps both. Human-gated.
Surface entity ↔ insight bridges (“this client maps to that insight you had”) for human consideration. Read-only; refuses to promote a record into an endorsed insight.
Writing and visuals (5)
Turn accumulated knowledge into output.
Long-form articles, blog posts, and Substack content from knowledge-base insights, with tone-of-voice and structure templates.
Extract your perspective on a topic (called by a content agent or by you).
Combine multiple insights into a coherent narrative.
Generate explanatory diagrams and infographics, iterating until the image is logically correct and the text clean. Uses Nano Banana (Gemini 2.5 Flash Image).
Generate images — infographics, diagrams, thumbnails, social graphics — with Nano Banana (Gemini 2.5 Flash Image).
Brain Dependency Graph (5)
Structural health of the directed dependency graph over your notes.
Full coherence sweep — staleness, lifecycle transitions, structural health, and a report.
Propagate staleness from a changed note — shows which downstream notes need review.
Lifecycle scores for insight and framework notes — detect which are crystallizing or becoming generative.
Detect productive contradictions — high semantic similarity with opposing conclusions — as synthesis opportunities.
Compare and selectively merge Brain directories across two agent instances: diff, pull, or bidirectional learn.
System and maintenance (10)
Keep the index, the history, and the agent healthy.
Orientation — introduces Cornelius, its core capabilities, main use cases, and the abilities marketplace.
Rebuild the Local Brain Search FAISS index to reflect vault changes.
Analyze knowledge-base structure and update the knowledge-base-analysis.md report.
Append a dated entry to Brain/CHANGELOG.md summarizing the session's knowledge-graph changes.
Stage, commit, and push with an approval gate.
Wrapper for scheduled playbooks — handles git sync before and after execution.
Update dashboard.yaml with current knowledge-base metrics for Trinity.
Verify that skills, commands, agents, and integrations are working.
Benchmark the Local Brain Search memory system with LLM-as-judge scoring.
Testing playbook for Local Brain Search memory improvements.
Shared contracts (6)
Rules other skills follow. Mostly not invoked directly — though scope-mount and reasoning-checks can be.
Which slice of the vault a read sees, decided by the read's role (voice / reasoning / lookup / perception-ingestion-maintenance), plus question-triggered reference mounts and a scope-grouped output format.
Four pre-conclusion checks — epistemic inversion (pre-mortem), attractor-state check, provenance weighting, reference-class anchor — run before a conclusion is finalized.
Distinguish research findings from hypotheses and speculative synthesis when creating notes from external sources.
The standard format for capturing and documenting insights in the knowledge base.
Techniques for extracting insights, probing your thinking, and conducting cognitive interviews.
Protocol for dated changelog files after significant agent sessions.