From Zero To Hero

A directed graph of prompts, models and tools with a hosted front end — where the vocabulary actually lives, why the canvas can lie to you, and the governance gap nobody mentions.

• 01Five primitives, one truth — the prompt text is the wire; dragging a connector just edits it for you
• 02Generate is the only node that costs you, and a drawn line on the canvas is not a guarantee the data still flows
• 03No public audit-log hook for Opal runs, and NotebookLM is now Gemini Notebook (16 Jul 2026) — same product, same URL

Vector RAG retrieves chunks, GraphRAG reasons over relationships — when that difference is worth the cost, and exactly which Azure services run it under real enterprise controls.

• 11The library is in maintenance mode — a reason to architect differently, not skip it
• 12Five use cases with a defensible "why vector fails," plus the one disqualifier: don't graph an FAQ
• 13Three Azure paths — self-hosted OSS, Agentic Retrieval (the default), or Fabric/Cosmos DB as a graph store

The model is the intelligence, the harness is what makes it useful — three disciplines that turn an LLM into an agent that actually ships.

• 21Agent = Model + Harness — a raw model can't hold state, execute code, or verify its own output, everything else is a harness-level feature
• 22Three nested disciplines: harness (what it can do) contains loop (when it stops or retries) contains graph (which paths are valid)
• 23Context rots as tokens pile up — compaction, context reset, tool-call offloading, and progressive disclosure beat raw context length

Anthropic's Figma-shaped agent · Chat on the left, canvas on the right — from turning it on to shipping to Claude Code.

• 31Design system first — skip it and every project is functional but generic
• 32Live code beats screenshots. Everything beats prompts
• 33/design-sync closes the loop with Claude Code — same account, same usage pool

Your vault becomes the memory, Claude Code becomes the agent — everything captured, cross-linked, ready to be acted on.

• 41The rule of the room: Agents read. Humans write — Claude never fabricates your notes
• 42Four replaceable layers: Markdown files → Git repo → Claude Code → Obsidian UI
• 43Obsidian 1.12 CLI: 54× faster than grep, 60%+ token reduction across skills

CLAUDE.md · Hooks · Skills · Subagents · Checkpoints · MCP — memory is engineered, not hoped for.

• 51Memory is a hierarchy of files you curate, not chat you hope survives
• 52Four layers: Enterprise · Project · User · Session
• 53Hooks + checkpoints + MCP servers = the difference between L2 and L4

One codebase, three models, three prices — routing beats picking. The right model matches the task's shape.

• 61Sonnet 5 $2/$10 · Opus 4.8 $5/$25 · Fable 5 $10/$50 per 1M tokens
• 62Start at Sonnet 5 — 82.1% SWE-bench for 40% of Opus's price
• 63~70% Sonnet · ~20% Opus · ~8% Fable 5 · ~2% Haiku — 4× lower bill

Not for lawyers. For engineers who will be asked "are we compliant?" and need to answer without lying.

• 71Aug 2, 2026 — EU AI Act is fully applicable. Not a drill
• 72Four frameworks. Different jobs. You need all four in some form
• 73Compliance is documentation you can't fake retroactively

Anthropic's canon, framework trade-offs, and the four anti-patterns that kill multi-agent projects.

• 81Most "multi-agent" systems should be a workflow, not agents
• 82The framework is a rounding error — the pattern is the design
• 83Production = HITL gates + observability + budgets

Less product tour, more value. Where it earns its credits, where it burns them, and when to reach for Claude instead.

• 91Manus doesn't answer — it goes and does
• 92The value is in delegation, not conversation
• 93Credits are the constraint — prompt like you're paying

A working engineer's briefing · One terminal, one loop, one job that just changed.

• 101The loop flipped — AI writes, you review
• 10230+ hours unattended on real code
• 103The skill is context, not prompts

Knowledge-work briefing · From "where did I save that" to "Claude already filed it."

• 111The unit changed — chat turns out, deliverables in
• 112The desktop is the surface, not the browser
• 113The skill is delegation — picking the right boundary

Patterns, not features — copy, paste, ship. Companion deck to the Code briefing.

• 121I · The terminal IS the IDE (patterns #1–5)
• 122II · Tests are the new prompt (#6–8)
• 123III · Context is a craft (#9–12)
• 124IV · Leverage compounds (#13–20)

Seven layers · One model · Zero trust — defend the model the way you defend the kernel.

• 131The attack surface is the agent's reach
• 132Guardrails are necessary — never sufficient
• 133Identity is the only durable boundary

Sixteen moves you can lift into a runbook before lunch. Companion to the Security briefing.

• 141I · Input boundary (4 patterns)
• 142II · Retrieval & data (4 patterns)
• 143III · Agent & tools (5 patterns)
• 144IV · Runtime & identity (3 patterns)

Cache · Route · Delegate · Plan — the $100 plan doesn't have a limit, your architecture does.

• 151Caching is the lever — all else amplifies it
• 152Opus is a tool, not a default
• 153Plan once, execute cheap

One API, 18 providers, 110+ models — with the IAM you already trust.

• 161Bedrock isn't a model — it's a control plane
• 162Six pillars. Five economic levers. Pick deliberately
• 163Identity & VPC are the moat, not the model

One ladder. Six topics. Three cloud tracks — pick the rung you're on, not the one you wish you were.

• 171The bottleneck isn't access to models — it's a missing rung
• 172Each rung has a build artifact, not a course completion
• 173Cloud track is the last call, not the first