Research catalog
Happy Overlord
Technologies, devices, and operating layers — each piece reviewable on its own before you combine them.
Cover story
Claude Opus 5.5: cheaper, faster, and four breaks
Opus 5.5 on Claude API: $4/$20, cache $0.20, four breaks — wiki to claude-api, claude-code, openclaw, litellm.
Technologies
Review each block alone
Special · Technology
mxbai-embed-large — Matryoshka at native 1024
Open Mixedbread embedder: full vector is 1024; truncate with dimensions / truncate_dim. Same method as OpenAI text-embedding-3-* (defaults 1536 / 3072) when you ask the API for 1024.
- RAGRetrieval pipelines that ground model answers in your own documents and indexes.
- AgentsAgent frameworks that plan, call tools, and run multi-step work for you.
- Vector DBsStores and search for embeddings — the memory layer behind retrieval and agents.
- WorkflowOrchestration for durable steps, retries, and long-running agent pipelines.
Folio
Devices that observe
Research
Vibe coding
Claude Code practice — loops, skills, hooks, MCP, and peers. Not a product page for vibe.codes.
Case studies
Write-ups, benchmarks, and builds that survived contact with reality.
Claude Code docs
The official product surface: install, agents, and reference docs.
CLAUDE.md · skills · hooks
How instructions, skills, hooks, and memory load into a session.
Context & prompting
What belongs in the window — and what you should leave out.
Loops & orchestration
Subagents, plan modes, and “until done” control loops.
MCP & tools
How tools attach, what sandboxes allow, and where MCP sits.
Origin & definition
Where the phrase came from — and what it actually claims.
Peers & IDEs
Cursor, Copilot, Aider, Codex — same job, different shells.
Quality & security
When vibe output needs tests, review, and hard security gates.
Team production
How teams actually ship with Claude Code day to day.
Research
Design × AI
Practice workflows and tools we run — plus Design resources (primary sources) linked into technologies and taxonomy.
Open practice
Workflows and tools you run — authored Design × AI practice by skill track.
Open resources
Primary sources for AI × design — citeable corpus linked from practice.
Case studies
Named papers and write-ups worth citing.
Design systems
Tokens, variables, and AI that lives inside the system.
End-to-end workflows
Brief → exploration → ship — full pipelines, not single prompts.
Image & motion
Still and motion generation as craft, not novelty demos.
Open-source tools
Open stacks you can run, fork, and audit yourself.
Origin & definition
Lock the vocabulary before picking tools or workflows.
Paid tools
Commercial products built for designers working with models.
Quality & ethics
IP, bias, accessibility, and how you evaluate the work.
Teams in production
Organizations already shipping design this way.
UX research
Synthesis, interviews, and testing when AI is in the loop.
Taxonomy
The vocabulary of this catalog
Layers, subcategories, device families, and OS lenses — each with a wiki page.
Layer
Agentic
Software that plans, calls tools, and runs multi-step work on behalf of a company — not raw model weights.
Layer
Infrastructure
Stores and graphs that keep embeddings, entities, and relationships available to agents and retrieval.
Layer
LLMs
Model serving and fine-tuning — how weights are run and adapted, not how agents are orchestrated.
Technology subcategories
- Agent Protocols
- Agentic Gateways
- Code Sandbox
- Coding Agents
- Evaluations
- Fine-tuning
- Frameworks
- Guardrails
- IAM
- Knowledge Graphs
- LLM Gateway
- Memory
- Observability
- RAG
- Runtime
- Search & Web Access
- Serving
- Structured Output
- Tool Integration
- Vector DBs
- Voice Agents
- Workflow Orchestration
Device families
Also reading
off-spine
OpenClaw when you already have Claude (or Claude Code)
Already on Claude Code? OpenClaw does not multiply your brain. Only a few pieces are unique: chat channels, always-on policy, and an optional local host.
C6
Closed execution loop vs recommendation dashboard
A dashboard that recommends is a colleague who points. Closing the loop needs an attempt, a result, a verified outcome, a feedback write.
Continued
Full archive- Stalled approval ≠ slow approvalOverdue is one threshold on one clock. Slow and stalled are different states with different remedies. No pe…
- HITL as a runtime primitive, not a FAQA week of reading approval chrome across twenty-four peers. All asserted, none enforced. A gate with fourte…
- Autonomy ↔ human authority: name both sidesTwenty pages say autonomy, twenty-seven say approval, and they barely overlap. Four screens of seven hundre…
- Approved baseline ≠ current timeline (the only honest slip)Fifteen of two hundred and twenty pages mention truth at all. Slip needs two objects: the approved baseline…