diff --git a/daily/2026-05-30-youtube.md b/daily/2026-05-30-youtube.md
new file mode 100644
index 0000000..6d81479
--- /dev/null
+++ b/daily/2026-05-30-youtube.md
@@ -0,0 +1,210 @@
+# YouTube Digest — 2026-05-30
+
+18 new videos across 40 channels. 11 worth full watch, 5 skim-only, 2 skipped.
+
+_All videos in this digest are marked `(transcript unavailable — summary based on title + description only)`. The transcript API returned `IpBlocked` from this environment's egress, and the yt-dlp fallback failed on `CERTIFICATE_VERIFY_FAILED` against youtube.com. Summaries lean on publisher descriptions and chapter markers, which were detailed for most videos. Durations marked `~` are estimated from the last chapter timestamp; `?` where no chapters were published._
+
+## Watch Fully (11)
+
+### [Reachy Mini: the $300 open source robot you can actually hack — Andres Marafioti, Hugging Face](https://www.youtube.com/watch?v=0jeZfjJMfmo) — AI Engineer (?min, category: ai-agents)
+**TL;DR:** Hugging Face's $300 hackable robot ships with a voice stack that hits sub-200ms time-to-first-audio after a deep TTS optimization pass. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- The TTS optimization writeup is the gold: Qwen3-TTS went from 0.8x real time → 5.8x real time by killing per-step CPU↔GPU round trips (500 autoregressive steps each), switching to a static KV cache, and capturing CUDA graphs. Direct playbook for any latency-sensitive Ara voice or streaming feature — the same three culprits show up in most "why is inference slow" investigations.
+- The voice loop is a reference architecture for low-latency agent UX: Parakeet transcription every 150ms with *partial results* feeding the LLM mid-sentence, Qwen 3.5 27B for reasoning, then the optimized TTS. The partial-results-as-context pattern is what makes the robot feel responsive — worth copying into any Ara voice loop.
+- Infra insight that scales beyond robots: at sub-200ms TTFA, network round trips match model latency, so they split LLM endpoints from "conversation node" endpoints in the load balancer to handle the different traffic shapes of talkative vs. quiet users. Familiars/Ara hit the same wall the moment voice is in the loop.
+- 7,500 units shipped unassembled at $300 — Hugging Face is using physical hardware as a distribution channel for agent UX research. Adjacent data point for Justin's "agentic product as a creative artifact" framing.
+
+**Tools mentioned:** Reachy Mini, Hugging Face, Qwen3-TTS, Parakeet (transcription), Qwen 3.5 27B, static KV cache, CUDA graphs
+
+**For /bookmark:** `https://www.youtube.com/watch?v=0jeZfjJMfmo`
+
+---
+
+### [Why your agents need decision traces, not just documents — Zach Blumenfeld, Neo4j](https://www.youtube.com/watch?v=B9h9ovW5H9U) — AI Engineer (?min, category: ai-agents)
+**TL;DR:** Neo4j extends RAG with a third layer — past decisions + reasoning embedded as vectors — so structurally similar prior cases surface alongside semantically similar docs. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- The three-layer context graph (short-term conversation, long-term entities, *reasoning traces as embedded vectors*) is the most directly applicable Ara memory architecture this week. The reasoning-trace layer is the one most agent stacks omit — and the one that lets a financial-analyst-style agent justify *why* it accepts or rejects, not just *what* it knows. Pairs with yesterday's "living memory is the moat" theme from the Accenture talk.
+- `uvx create-context-graph` is a one-command scaffold that gives backend + frontend + demo data + MCP server out of the box, with 22 prebuilt domains or LLM-generated ontologies. This is the fastest possible way for Justin to prototype a context-graph memory layer for Ara without standing up Neo4j from scratch.
+- The `neo4j-agent-memory` package does entity extraction through a spaCy → GLiNER → LLM pipeline with deduplication and merging baked in — the dedupe step is the one most "naive memory" projects regret skipping. Plugs directly into pydantic AI, LangGraph, Crew, and Google ADK.
+- The MCP-server-by-default packaging is a quiet signal: agent infra is increasingly shipping with MCP as the integration story rather than custom SDKs. Worth noting as Justin designs Ara's external interface.
+
+**Tools mentioned:** Neo4j, `uvx create-context-graph`, `neo4j-agent-memory`, spaCy, GLiNER, MCP, pydantic AI, LangGraph, Crew, Google ADK
+
+**For /bookmark:** `https://www.youtube.com/watch?v=B9h9ovW5H9U`
+
+---
+
+### [Reverse engineering a Viking VOIP phone protocol with Claude Code — Boris Starkov, Eleven Labs](https://www.youtube.com/watch?v=V-L0INGTEOg) — AI Engineer (?min, category: ai-agents)
+**TL;DR:** ElevenLabs cracked an undocumented Windows-XP-era phone protocol in days with Claude Code orchestrating the brute force, packet inspection, and checksum derivation. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- The agent loop is the lesson: Claude brute-forced all 676 two-letter command combos to find 80 valid ones, set up a TCP proxy between a Windows VM and the phone to log the proprietary software's traffic, then derived a one-byte checksum formula by feeding known input/output pairs through a closed loop. Three different reverse-engineering techniques chained by a single agent — a template for any "the docs don't exist, figure it out" Ara task.
+- Boris's framing — *"Claude orchestrated, I was the hands"* — is the cleanest articulation yet of the human-as-actuator pattern: the LLM plans and reasons, the human does the physical-world actions (reboot, count beeps) and reports back. Important framing for any Ara/Familiars task that crosses the digital/physical boundary.
+- The deliverable was packaged as a Claude Code *skill* (the protocol is open-sourced) — concrete reuse pattern: investigations should output a portable skill, not just a fix. Maps directly onto Justin's `.claude/skills/` discipline.
+- The end-state demo is a creative-engineering flex worth stealing the vibe of: a red phone booth in the AI Engineer Europe lobby where the receiver connects to a Michael Caine voice agent quizzing you on British AI history. Strong "agent + artifact" reference for Frontier Colony's storytelling.
+
+**Tools mentioned:** Claude Code, Claude Code skills, TCP proxy, Windows VM, ElevenLabs voice agent
+
+**For /bookmark:** `https://www.youtube.com/watch?v=V-L0INGTEOg`
+
+---
+
+### [New Claude Opus 4.8: 15 Things You May've Missed](https://www.youtube.com/watch?v=aJvP3nXWkwM) — AI Explained (~21min, category: ai-agents)
+**TL;DR:** Philip mines the 244-page Opus 4.8 system card for 15 details from honesty calibration to model welfare to the dynamic-workflows release. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- This is the *system-card-grounded* Opus 4.8 video; Nate Herk's WATCH below is the workflow-grounded one. Watch both. Highlights to listen for: "Honesty?" (2:26) and "Flagging Uncertainty" (4:37) — calibration improvements that change how much anti-hallucination scaffolding Ara actually needs. "Misalignment Concerns" (13:10) and "Meta Inabilities" (16:22) are the safety/limitations sections most reaction videos skip.
+- Chapter "Fast Mode" (18:50) + "Dynamic Workflows" (20:21) line up with the Theo video below and the linked `claude.com/blog/introducing-…` post. Dynamic workflows in Claude Code is the new release surface to dig into — directly relevant to how Justin structures Ara's prompt flows.
+- "Mythos in Weeks" (00:49) and "Mythos will be even better" (8:54) imply a near-term successor codenamed Mythos — useful planning signal for any model-version commitments in Ara.
+- Links the GDPVal economic-utility eval (artificialanalysis.ai/evaluations/omniscience + arXiv 2510.04374) and Dario's "technical debt" talk — both are background reading worth queuing.
+
+**Tools mentioned:** Claude Opus 4.8, Anthropic system card, Claude Code Dynamic Workflows, Fast Mode, GDPVal
+
+**For /bookmark:** `https://www.youtube.com/watch?v=aJvP3nXWkwM`
+
+---
+
+### [I Turned Claude Opus 4.8 Into My Entire AI Operating System](https://www.youtube.com/watch?v=0WDkwMxj13s) — Nate Herk (~29min, category: automation)
+**TL;DR:** A start-to-finish framework — Context, Connections, Capabilities, Cadence — for running an entire business out of Claude Code as the default surface. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- The **Four C's** (Context Is King 4:46 → Connections 8:20 → Skills/Capabilities 20:12 → Cadence) is the most useful taxonomy this week for organizing an AI-OS-style setup. Justin already has the Obsidian-vault context layer; the Connections + Skills chapters are where there's the most leverage to copy.
+- "One Source of Truth" (14:32) is the spec for the file-organization discipline that makes Claude Code reliable across sessions — directly applicable to how `~/Documents/Obsidian Vault` and `ai-research-vault` should be structured to avoid drift.
+- The **Bike Method** (15:24) is the framing for graduating agents from training wheels to autonomy by progressively removing supervision on small wins — a more memorable version of the shadow → advisory → controlled-autonomy ladder from yesterday's Accenture talk. Use the Bike Method language internally; it's stickier.
+- "Do You Need a Dashboard?" (25:20) is a worthwhile contrarian segment given Justin's design instincts — the argument is usually *no*, because the terminal/markdown surface beats most custom dashboards. Counter-evidence worth weighing before building any Ara dashboard.
+
+**Tools mentioned:** Claude Code, Claude Opus 4.8, Skool community resources, Glaido (voice-to-text), Hostinger VPS
+
+**For /bookmark:** `https://www.youtube.com/watch?v=0WDkwMxj13s`
+
+---
+
+### [Anthropic fights back](https://www.youtube.com/watch?v=_goOUJkkxUk) — Theo - t3.gg (?min, category: web-dev)
+**TL;DR:** Theo reacts to the Opus 4.8 launch *and* the "ultracode" release — both quietly significant moves that frame Anthropic's positioning against OpenAI/Cursor. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- Theo's reactions are the fastest way to triangulate Anthropic's positioning against the prevailing web-dev/Cursor crowd narrative. As someone building for an Anthropic portfolio fit, Justin should know what the loudest competing influencer is saying about each release within 24h — even when (especially when) it disagrees with the AI Explained framing.
+- "Ultracode" is the new surface to investigate — it's mentioned in passing across multiple sources this week but underexplained. This video is a low-cost way to learn what it is and whether it overlaps with anything Ara is doing.
+- Useful triangulation against the AI Explained Opus 4.8 deep-dive above — same release, very different framing. The delta between "system card details" and "developer-tooling reaction" is the most informative read on how the release will actually land.
+
+**Tools mentioned:** Claude Opus 4.8, "ultracode", CodeRabbit
+
+**For /bookmark:** `https://www.youtube.com/watch?v=_goOUJkkxUk`
+
+---
+
+### [How to use Obsidian with Claude in 61 seconds](https://www.youtube.com/shorts/2VSmDS3ALnw) — Greg Isenberg (~1min, category: indie-startup)
+**TL;DR:** A 60-second teaser for a long-form sit-down with Internet Vin on using Obsidian + Claude Code as a personal thinking partner with custom slash commands. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- *This is the highest-relevance video in the digest for Justin's actual workflow.* Vin's setup mirrors Justin's almost exactly: Obsidian vault of interlinked markdown + Claude Code + the new Obsidian CLI that lets Claude see *file relationships*, not just file contents. The Short is the teaser — the *real artifact* to chase is the longer pod and the linked skills page: `https://startup-ideas-pod.link/obsidian-commands`.
+- The custom slash commands described — `/trace`, `/connect`, `/ideas`, `/ghost`, `/drift`, `/challenge` — are a template for Justin's own `.claude/skills` library. Each one names a specific *thinking move* on top of the vault: trace idea evolution, find connections, generate ideas, surface contradictions, detect drift, steelman the opposite. Directly portable; copy the naming, then adapt to your vault's structure.
+- The Obsidian CLI angle is the new piece — even if Justin already pipes vault files into Claude, the CLI exposing the *link graph* changes what queries are possible (e.g. "find every note that descended from this 2023 seed idea").
+- Per the Shorts rule this would default to SKIM, but the description is essentially the long-form description and the topic is so directly aligned with Justin's existing vault setup that the longer pod is worth bookmarking now.
+
+**Tools mentioned:** Claude Code, Obsidian, Obsidian CLI, custom slash commands (`/trace`, `/connect`, `/ideas`, `/ghost`, `/drift`, `/challenge`)
+
+**For /bookmark:** `https://www.youtube.com/shorts/2VSmDS3ALnw`
+
+---
+
+### [Every Hermes Concept explained for Normal People](https://www.youtube.com/watch?v=L0tQrTQBmjI) — Jack (~41min, category: indie-startup)
+**TL;DR:** A 21-concept tour of the Hermes AI agent platform — from `soul.md` and sub-agents to MCPs, heartbeats/cron, and the "operating system" framing. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- Hermes is parallel to Justin's own AI-OS vision (Familiars + Ara as a personal operating system); studying a different shipping product's vocabulary and architecture is the cheapest competitive intel. Specific concepts worth comparing 1:1: "One Brain 22 Mouths" (4:14), "soul.md" (18:17), "Spin Up Sub Agents" (33:53), "Heartbeat And Cron Jobs" (35:35), "Goals vs Super Goals" (32:49).
+- "MCPs Simply Explained" (23:13) + "Connect Hermes To Claude" (40:31) imply Hermes is positioning itself as an MCP-native agent host with first-class Claude integration. Pattern to study: how does Hermes UX the user's MCP servers? Apply learnings to how Ara should surface its own integrations.
+- "Safety And Least Access" (31:49) and "Build Hermes Muscle Memory" (25:36) are the segments that get nearest to the *daily-use* questions that any AI-OS product has to answer; Hermes's answers will be useful comparison points for Ara's onboarding.
+- "Run It Locally Private" (14:20) is worth noting as a differentiator — Justin's products will need a clear stance on local vs. cloud, and seeing Hermes's framing helps inform the pitch.
+
+**Tools mentioned:** Hermes, ChatGPT, Claude, GitHub, Pinecone, MCPs, Zapier, Glaido
+
+**For /bookmark:** `https://www.youtube.com/watch?v=L0tQrTQBmjI`
+
+---
+
+### [This 100% open-source terminal is insane… just watch](https://www.youtube.com/watch?v=8jDXI4_rJOE) — David Ondrej (?min, category: ai-agents)
+**TL;DR:** Walkthrough of CMUX, a fully open-source AI-agent terminal — a hackable surface for running coding agents without Claude Code or Cursor lock-in. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- CMUX (`cmux.com`) is worth a 10-minute evaluation as a *reference implementation* even if Justin stays on Claude Code — an open-source agent terminal exposes the design decisions that closed tools hide (session state, tool routing, sandbox boundaries). Compare its UX against Claude Code's defaults to identify what feels good and what doesn't.
+- For any agent feature Justin builds that needs to live *outside* Anthropic's closed surfaces — local-only flows, white-labeled embeddings, anything the portfolio review would want to see decoupled from a single vendor — an open-source baseline like CMUX is the right thing to fork or learn from rather than reinvent.
+- Ondrej's hype-frame is heavy, so weight the demo over the editorial; but he generally surfaces tools weeks before they hit the broader feed.
+
+**Tools mentioned:** CMUX (`cmux.com`), Scale Software ("we're hiring" — Ondrej-affiliated), Skool / New Society community
+
+**For /bookmark:** `https://www.youtube.com/watch?v=8jDXI4_rJOE`
+
+---
+
+### [Import & Export | FLORA AI](https://www.youtube.com/watch?v=RTFip5CZDBw) — FLORA (~2min, category: competitive-intel)
+**TL;DR:** FLORA now connects directly to Google Drive — authenticate once, then import → run workflow → export back without leaving the canvas. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- The competitive read for Nebula Nodes: FLORA is investing in the *handoff edges* of the workflow (where assets enter and leave the canvas), not just the nodes themselves. The Google Drive integration removes a download/upload round-trip that's the most-cited friction point for canvas-style creative tools. If Nebula Nodes targets the same buyer, parity here is table stakes; the differentiated play is the *next* integration (Figma? Dropbox? Notion?).
+- Chapter breakdown (Authenticate → Import → Build workflow → Export → Disconnect) is the canonical onboarding flow for any creative integration. Worth replicating section-for-section as the spec for Nebula Nodes' first integration.
+- Docs link `https://docs.flora.ai/nodes/export-node` is a direct competitive-intel artifact — read FLORA's node docs for naming conventions, granularity, and what they *don't* expose. Their public docs surface is itself a competitive signal.
+
+**Tools mentioned:** FLORA AI, Google Drive (OAuth integration), Export Node
+
+**For /bookmark:** `https://www.youtube.com/watch?v=RTFip5CZDBw`
+
+---
+
+### [How to combine Batch Nodes with a Technique](https://www.youtube.com/watch?v=pbNebYZlOKY) — FLORA (?min, category: competitive-intel)
+**TL;DR:** FLORA pairs Batch Nodes with reusable Techniques (saved brand/campaign styles) so a single connection runs your house style across an entire catalog in parallel. _(transcript unavailable — summary based on title + description only)_
+
+**Key insights:**
+- "Techniques as packaged recipes" is the noun Nebula Nodes needs to wrestle with. FLORA has named the unit — a reusable style preset that anyone on the team can apply — and that naming is a meaningful competitive moat. Either adopt the same vocabulary (and compete on execution) or find a better noun.
+- The Batch × Technique product combo is the *team* play, not the individual-creator play: it only matters once a brand-look needs to be reapplied across catalogs by people who didn't create the look. This signals FLORA's wedge toward in-house creative teams and agencies, where Nebula Nodes will eventually meet them.
+- The parallel-execution framing ("returns a full set of on-style outputs that flow straight into the rest of your workflow") is a UX bar to match: batch shouldn't break the canvas metaphor — outputs should land back as nodes that downstream steps can consume. Worth designing for from day one.
+
+**Tools mentioned:** FLORA AI, Batch Nodes, Techniques
+
+**For /bookmark:** `https://www.youtube.com/watch?v=pbNebYZlOKY`
+
+---
+
+## Skim Summary (5)
+
+### [Treat AI Like a Junior Engineer](https://www.youtube.com/shorts/tUZvIPRVlXA) — Vercel (~1min, Short)
+A clip of Guillermo Rauch (Vercel CEO) framing prompt engineering as people management: when an agent goes off the rails, the highest-leverage move is to "reset" the context rather than keep correcting in place. Logged because it crystallizes a useful instinct Justin already has — but the source is a 1-minute promo for a longer "Show Me Your Stack" interview on @boysclubworld, which is where the actual depth lives. _(transcript unavailable — summary based on title + description only)_
+
+**Takeaway:** Make "context reset" an explicit first-class action in any agent loop, not a last resort — then watch the long-form Boys Club interview for Rauch's full framing.
+
+---
+
+### [Self Improving AI actually solves everything](https://www.youtube.com/shorts/IYN2jfJHJd4) — Matthew Berman (~1min, Short)
+A no-substance hype Short pitching a sponsor ("Pioneer") and the self-improving-AI thesis without any specific paper, model, or result referenced in the description. Logged per the Shorts rule but offers no actionable content beyond the title's framing. _(transcript unavailable — summary based on title + description only)_
+
+**Takeaway:** Nothing to act on; if self-improving-AI claims become concrete in a later Berman long-form, pick it up there.
+
+---
+
+### [The forgotten developer who saved JavaScript...](https://www.youtube.com/watch?v=JfPWbttemYE) — Fireship (?min, category: web-dev)
+A typical Fireship history piece — likely on the person/event that rescued JavaScript at a pivotal moment (Brendan Eich's 10-day-language origin, Crockford's "good parts," or the V8/Node era are the usual candidates). Entertaining and good for general engineering literacy; no specific tools, frameworks, or releases hinted in the description beyond a Clerk CLI sponsor mention. _(transcript unavailable — summary based on title + description only)_
+
+**Takeaway:** Skim if you want the trivia and the historical context; nothing operational for Ara/Familiars/Nebula Nodes.
+
+---
+
+### [Gemini co-leads on project origins and what's next](https://www.youtube.com/watch?v=8hfpLa5wPGo) — Google Developers (?min, category: platform)
+Logan Kilpatrick interviews Jeff Dean, Koray Kavukcuoglu, Noam Shazeer, and Oriol Vinyals at Gradient Canopy on the launch of Gemini 3.5 Flash. The pitched themes are the origin story of the Gemini project, the single-unified-model bet, why each Flash generation now beats the previous Pro, and where they're going next. Useful as competitive-intel sizing — Gemini 3.5 Flash is the model Cole Medin's WATCH entry from yesterday wires into the design step of a multi-provider workflow. _(transcript unavailable — summary based on title + description only)_
+
+**Takeaway:** Skim for two things — (1) the specific Flash-beats-previous-Pro claim and what they attribute it to, and (2) any roadmap hints about the next Pro generation that would affect provider-routing decisions.
+
+---
+
+### [Entrepreneurship isn't what you think it is](https://www.youtube.com/shorts/ZO9sYrw7i6c) — David Ondrej (~1min, Short)
+A no-description motivational Short on entrepreneurship, almost certainly a teaser/clip from a longer Ondrej video. Generic enough that there's nothing concrete to extract; logged because the Shorts rule says SKIM unless clearly off-thesis. _(transcript unavailable — summary based on title + description only)_
+
+**Takeaway:** Nothing actionable — skip unless it surfaces as the trailer for a longer founder/portfolio piece that's worth the full watch.
+
+---
+
+## Skip (2)
+
+- [Were Neanderthals Culturally Modern Humans? - David Reich](https://www.youtube.com/shorts/ci7_eJTrGhQ) — Dwarkesh Patel — _Reason: Anthropology/genetics Short clipped from a longer David Reich interview. Genuinely interesting but fully off-thesis (no AI, agent, or design-engineering hook). Override if you want the long-form Reich interview; chase that one in Dwarkesh's main feed, not this clip._
+- [The Pope is into AI](https://www.youtube.com/watch?v=HoMvCjnpAJ8) — Matthew Berman — _Reason: News reaction to Pope Leo XIV's "Magnifica Humanitas" AI encyclical (the Vatican link is in the description). Real cultural data point for how the church frames AI, but no technical, agentic, or product content. Override if you specifically want the religious/ethical framing — the Vatican link itself is the primary source._