Skip to content

Guides and resources for AI engineers: an agents course, five handbooks, a hands-on second-brain project and a vetted reading list.

12.7 KB
Raw Blame
{
  "_about": "Terms shown as nodes on the home graph. Each term links to the pages that mention it (any alias, case-insensitive unless \"case\": true). Edit freely, then rerun tools/build_graph_site.py.",
  "terms": [
    {"id": "second-brain", "label": "Second brain", "aliases": ["second brain", "second-brain"], "def": "One local folder of plain text that an agent reads and writes, holding what you have read and understood, that you can ask questions of."},
    {"id": "llm-wiki", "label": "LLM wiki", "aliases": ["llm wiki", "llm-wiki"], "def": "Karpathy's pattern: the model compiles sources into a linked wiki once, instead of rediscovering them from chunks on every question."},
    {"id": "vault", "label": "Vault", "aliases": ["vault"], "def": "The folder of markdown the agent maintains. Obsidian calls it a vault; to the agent it is just files."},
    {"id": "obsidian", "label": "Obsidian", "aliases": ["obsidian"], "def": "The markdown editor used to read the vault, browse its graph and clip sources. The agent never needs it running."},
    {"id": "claude-code", "label": "Claude Code", "aliases": ["claude code"], "def": "The agent that does the filing: reads raw sources, writes and links pages, runs the maintenance."},
    {"id": "claude-md", "label": "CLAUDE.md", "aliases": ["CLAUDE.md"], "case": true, "def": "Standing instructions the agent reads at the start of every session: who you are, the page contracts, the vault's rules."},
    {"id": "mcp", "label": "MCP", "aliases": ["MCP", "Model Context Protocol"], "case": true, "def": "Model Context Protocol: the standard plug that lets an agent reach tools and data, such as your vault from anywhere."},
    {"id": "raw", "label": "raw/", "aliases": ["raw/", "raw layer", "raw sources"], "def": "The inbox and archive: clipped articles, transcripts, PDFs. Written once, never edited, rarely read again."},
    {"id": "two-layers", "label": "Two layers", "aliases": ["two layers", "project layer", "knowledge layer"], "def": "A wiki for what you know and a project layer for what you are doing, kept apart so each stays useful."},
    {"id": "ingest", "label": "Ingest", "aliases": ["ingest", "ingests", "ingested", "ingesting", "ingestion"], "def": "Turning one raw source into linked wiki pages. Not finished until the new pages are connected to what was already there."},
    {"id": "wikilink", "label": "Wikilink", "aliases": ["wikilink", "wikilinks", "[["], "def": "A [[Page name]] link between notes. The cheapest possible graph edge, and the reason the vault compounds."},
    {"id": "backlinks", "label": "Backlinks", "aliases": ["backlink", "backlinks", "inbound link", "inbound links"], "def": "The links pointing at a page. How you find everything that leans on an idea without searching for it."},
    {"id": "atomic-pages", "label": "Atomic pages", "aliases": ["atomic page", "atomic pages", "one idea per page", "atomic"], "def": "One idea per page, so a link means one thing and a page can be reused across every context it belongs to."},
    {"id": "page-types", "label": "Page types", "aliases": ["page type", "page types", "concept page", "entity page", "source page", "synthesis page"], "def": "Source, concept, entity, synthesis: each with its own contract for what goes in it and how it links."},
    {"id": "frontmatter", "label": "Frontmatter", "aliases": ["frontmatter", "front matter", "yaml"], "def": "The YAML block at the top of a page: type, sources, dates, status. What makes the vault queryable without a database."},
    {"id": "aliases", "label": "Aliases", "aliases": ["alias", "aliases"], "def": "Other names a page answers to, so links and searches land on one page instead of spawning duplicates."},
    {"id": "contradictions", "label": "Contradictions", "aliases": ["contradiction", "contradictions", "contradicts", "supersession", "superseded"], "def": "When a new source disagrees, both positions stay, dated and sourced. The record of what you believed and why."},
    {"id": "index-log", "label": "index.md and log.md", "aliases": ["index.md", "log.md"], "case": true, "def": "The catalog the agent reads first, and one line per run so every change to the vault is auditable."},
    {"id": "knowledge-graph", "label": "Knowledge graph", "aliases": ["knowledge graph", "knowledge graphs", "graph view"], "def": "Pages as nodes, links as edges. In this setup the graph is the filesystem: no server, no separate store."},
    {"id": "typed-links", "label": "Typed links", "aliases": ["typed link", "typed links", "typed relationship"], "def": "A link that says how two pages relate, such as supports, contradicts or extends. Only typed when the source says so."},
    {"id": "orphans", "label": "Orphans", "aliases": ["orphan", "orphans", "orphan rate", "unlinked"], "def": "Pages nothing links to. Invisible within a week, and the first number to watch when a vault is decaying."},
    {"id": "hubs", "label": "Hubs and bridges", "aliases": ["hub", "hubs", "bridge", "bridges", "betweenness"], "def": "Hubs are pages everything links to; bridges connect clusters that would otherwise not touch. Both tell you where the thinking is."},
    {"id": "clusters", "label": "Clusters", "aliases": ["cluster", "clusters", "community", "communities"], "def": "Groups of densely linked pages. The ones you did not plan are where the vault is telling you something."},
    {"id": "graphrag", "label": "GraphRAG", "aliases": ["graphrag", "graph rag"], "def": "Retrieval over an extracted entity graph with community summaries. Built for global questions about a whole corpus."},
    {"id": "rag", "label": "RAG", "aliases": ["RAG", "retrieval-augmented", "retrieval augmented"], "case": true, "def": "Retrieval-augmented generation: fetch chunks by similarity at question time. Nothing accumulates between questions."},
    {"id": "vectors", "label": "Embeddings and vectors", "aliases": ["embedding", "embeddings", "vector database", "vector store", "vectors", "vector search"], "def": "Similarity search over chunk embeddings. Worth adding only once a structured, linked vault stops answering."},
    {"id": "context-window", "label": "Context window", "aliases": ["context window", "context budget", "context windows"], "def": "What the model can see at once. Everything the agent knows in a session competes for this budget."},
    {"id": "lost-in-middle", "label": "Lost in the middle", "aliases": ["lost in the middle"], "def": "Models recall the start and end of a long context far better than the middle. Why stuffing everything in fails."},
    {"id": "skills", "label": "Skills", "aliases": ["skill", "skills", "SKILL.md"], "def": "A folder with a SKILL.md: instructions the agent loads only when the task matches. One per workflow."},
    {"id": "slash-commands", "label": "Slash commands", "aliases": ["slash command", "slash commands", "/ingest", "/lint", "/ask"], "def": "Short, scoped entry points into a skill, typed as /name. The handle you actually use day to day."},
    {"id": "subagents", "label": "Subagents", "aliases": ["subagent", "subagents", "sub-agent"], "def": "Separate agents with their own context and a narrow tool list. Most of them here are read-only by design."},
    {"id": "hooks", "label": "Hooks", "aliases": ["hook", "hooks", "PreToolUse", "PreCompact"], "def": "Code the harness runs on events, before a tool call or after an edit. Rules the model cannot forget."},
    {"id": "scheduled", "label": "Scheduled maintenance", "aliases": ["schedule", "scheduled", "cron", "nightly", "overnight"], "def": "Lint, link and review runs on a timer, so the vault files itself while you sleep."},
    {"id": "lint", "label": "Lint and health", "aliases": ["lint", "linting", "linter", "health check"], "def": "Mechanical checks for broken links, orphans, missing frontmatter and stale pages. The vault's test suite."},
    {"id": "git", "label": "Git", "aliases": ["git", "commit", "commits", "rollback"], "def": "Version history for the vault. Every agent run becomes reviewable and every bad run reversible."},
    {"id": "privacy", "label": "Privacy and secrets", "aliases": ["privacy", "secret", "secrets", "api key", "api keys"], "def": "What must never reach a page, a prompt or a public export, and the checks that keep it out."},
    {"id": "web-clipper", "label": "Web Clipper", "aliases": ["web clipper", "clipper", "clip", "clipped"], "def": "Obsidian's browser extension. One click turns an article into a markdown file in raw/."},
    {"id": "transcripts", "label": "Transcripts", "aliases": ["transcript", "transcripts", "transcription"], "def": "Video, voice and meetings turned into text so they can be ingested like anything else."},
    {"id": "agent", "label": "Agent", "aliases": ["agent", "agents", "agentic"], "def": "A model in a loop with tools, deciding its own next step until a goal is met or a stop rule fires."},
    {"id": "five-layers", "label": "Five layers", "aliases": ["five layers", "five-layer"], "def": "The course's frame for everything around the model: context, loop, gate, harness, evals."},
    {"id": "workflows", "label": "Workflows vs agents", "aliases": ["workflow", "workflows"], "def": "Code decides the next step in a workflow; the model decides in an agent. Most production systems are a hybrid."},
    {"id": "context-engineering", "label": "Context engineering", "aliases": ["context engineering", "four places", "the four places"], "def": "Deciding what goes into the system prompt, tools, history and tail, and in what order, so the model sees the right things."},
    {"id": "caching", "label": "Prompt caching", "aliases": ["prompt caching", "cache", "caching", "cached"], "def": "Reusing the stable prefix of a prompt across calls. Cheaper and faster, if the context is laid out for it."},
    {"id": "memory", "label": "Memory", "aliases": ["memory", "memories"], "def": "What persists between sessions. Here it is files the agent reads back, not a hidden feature of the model."},
    {"id": "loop", "label": "Agent loop", "aliases": ["agent loop", "loop", "loops", "ralph"], "def": "Goal, act, check, repeat. The loop is the control system around the model, and most failures live in it."},
    {"id": "stop-conditions", "label": "Stop conditions", "aliases": ["stop condition", "stop conditions", "stop rule", "budget"], "def": "When a loop must end: goal met, budget spent, no progress. A loop without one runs until it does damage."},
    {"id": "critics", "label": "Critics and checkers", "aliases": ["critic", "critics", "checker", "checkers", "verifier", "verification"], "def": "A check outside the model that judges the result: tests, a compiler, a separate reviewer. Self-review is not one."},
    {"id": "gate", "label": "The gate", "aliases": ["gate", "gates", "classifier", "classifiers", "routing", "router"], "def": "A cheap, narrow decision in front of an expensive model: one bounded question per item, route on the answer."},
    {"id": "system-one", "label": "System One models", "aliases": ["system one", "jev"], "def": "Small fast models that return typed decisions with confidence instead of text. Jev is one way to build the gate."},
    {"id": "harness", "label": "Harness", "aliases": ["harness", "harnesses"], "def": "The office around the model: containment, guides, sensors and permissions, built from the outside in."},
    {"id": "guardrails", "label": "Guardrails and permissions", "aliases": ["guardrail", "guardrails", "permission", "permissions", "sandbox", "containment", "dry-run", "dry run"], "def": "What the agent is allowed to touch, what it must ask about, and the walls that hold when it is wrong."},
    {"id": "tools-tool-use", "label": "Tool use", "aliases": ["tool use", "tool_use", "tool call", "tool calls"], "def": "The model asking the harness to run something on its behalf. Every capability an agent has arrives this way."},
    {"id": "evals", "label": "Evals", "aliases": ["eval", "evals", "evaluation", "evaluations"], "def": "The same test, every month: checks on real behaviour so a prompt or model change cannot quietly break things."},
    {"id": "golden-set", "label": "Golden set", "aliases": ["golden set", "golden sets", "golden.jsonl"], "def": "The fixed collection of cases every eval runs against, including the failures you have already seen."},
    {"id": "llm-judge", "label": "LLM as judge", "aliases": ["llm-as-judge", "llm as judge", "judge", "judges"], "def": "A model grading outputs against a rubric. Useful once the judge itself has been checked against human labels."},
    {"id": "traces", "label": "Traces", "aliases": ["trace", "traces", "tracing", "trajectory", "trajectories"], "def": "The full record of what an agent did, step by step. What behavioural evals and debugging both read."},
    {"id": "production", "label": "Production", "aliases": ["production", "deployed", "gateway", "gateways"], "def": "The step from a demo that works once to a system with tracing, cost control, security and an owner."}
  ]
}