PowerUp AI Mastermind — June 25, 2026
“Boris Cherny — the Claude Code developer — says he doesn’t write code anymore. He writes loops. And you can have loops within loops.” — Lou
This Week in 30 Seconds
- The hub is now one front door — Lou consolidated the gallery, the AI query engine, the here.now site and the hub into a single Astro site at aimmhub.coachlou.com: recaps, replays, collections, skills, insights, plus a built-in “Ask AI” running on Groq + OpenAI’s free 120B model. Magic-link login, dark/light themes. He showed two designs (Claude-style sidebar vs. a minimalist “Google” look); the group leaned toward the Claude-style sidebar.
- Claude Design — handy, not essential — Lou’s verdict after two builds: the markup-the-screen interactivity and Figma↔Code transport are nice, but token burn is brutal (two or three iterations exhausts the five-hour limit on a $20 plan), and for non-designers the output isn’t meaningfully better than asking 4.8 directly. Don’s tip: register your style sheet with Design, then drive it from cheaper surfaces.
- The offload pattern — Donald runs GLM 5.2 (via OpenRouter) as a cheap second worker: Opus plans, GLM does the heavy lifting in another terminal, and he never hits a limit. Lou’s caveat: GLM is 2–3× slower, API cost compounds (~$1k/mo at full-time use), and OpenRouter has no guaranteed uptime — so treat it as a backup/offload, not a Claude replacement.
- Scott’s long-tail win — Scott characterized 90M HPC jobs by exploring an unfamiliar database with Claude and generating Python that answered a question his CTO had chased for months. The official tooling will never build “Scott’s weird query” — but he can.
- Don’t rent autonomy — replicate it — Dirk balked at Hermes’ stacked cost, asked Claude to make Claude Code behave like it, and got a four-hour hands-off run. Lou’s teaching: that autonomy is a loop in the harness, not a product you must buy.
- Write loops, not code — Lou generalized the moment into the session’s spine: define a goal, a rubric, and a stop condition, then let the model iterate — and nest loops within loops. Effort tunes persistence, not intelligence.
- Grill before you build — Lou’s recommendation of the week: Matt Pocock (GitHub/YouTube) and his “grilling” skill — exhaustive depth-first interrogation of blind spots before you act, feeding a full PRD → GitHub-issues → triage → PR → merge pipeline. Lou’s adaptation: swap the PRD for a content brief and feed your writing team.
- The human stuff — a genuinely hard week for Lou: father-in-law’s cancer returned (family flew back to Toronto for care), his mom’s 90th birthday and a nephew’s wedding celebration all landing at once. As Bally put it: “that was your quiet week, Lou.”
1 — Claude Design: Handy, Not Essential
Lou opened with a candid review of Claude’s Design feature after building two versions of the knowledge site through it. What he liked: you can mark up the rendered screen and the comment flows straight into the chat; you can edit text and move elements live; and it connects to Figma and exports to Claude Code, so designers used to that pipeline will get real value. What gave him pause: the token burn. “Two, maybe three significant iterations and the five-hour limit is gone on a $20 plan.” His bottom line for non-designers: “the website it created wasn’t that much different than the one 4.8 created on its own — unless you’re doing something really fancy, I’m not sure it’s an essential upgrade.”
Don Back offered the cost-saver: load your style sheet / corporate style into Design and register it, then use the cheaper chat surfaces, which can reach back to that design sheet — “you can go halfway, and certainly less expensive than going all the way through Design.” Lou’s own direction is the opposite of design-heavy: he’s deliberately building chat-based apps with no UI at all.
💡 What This Means for You
If you’re not a designer, you probably already have what you need in Opus 4.8 plus the front-end design skill. Reserve Claude Design for genuinely visual, Figma-centric work — and watch the token meter.
2 — One Front Door: The Unified Knowledge Hub
The centerpiece walkthrough. Lou has folded everything that used to live in separate places — the gallery, the perplexity-style query engine, the here.now site, the old hub — “under the one hood” at aimmhub.coachlou.com. The flow that built it: design the look in Claude Design, export it, hand it to Claude Code, and have it generate an Astro site (pre-rendered HTML so Cloudflare serves it fast). It now carries recaps, replay videos, collections (the teaching blocks), skills, and insights, with magic-link login (no passwords) and automatic dark/light theming.
The piece members reacted to most was the integrated “Ask AI” — query the whole knowledge base in natural language, get an answer plus the sources it pulled from, running on the Groq interface with OpenAI’s free 120B open model (fast, non-streaming, “free still”). “You don’t have to point your Claude at my knowledge base — it’s already built into this.”
Two design notes surfaced from the group: a couple of articles were missing the next/previous backlink (Lou: that happens when you’re on a “leaf entity” in the graph — he’ll add next/prev navigation), and a debate over the two looks. Donald liked the Apple-ish icons and tags of the Claude-style version; Bally and Don leaned minimalist, then the room circled back to “lean to the Claude one.” Lou will likely keep the sidebar version live and keep the minimalist one in a work tree.
A quieter but important point: the maintenance tax that used to kill second-brain systems is gone. Lou: in old Obsidian you had to physically type the links between articles — “it became administratively prohibitive to maintain. Now with the wiki, all of those links are automatically established by the logic of the AI.” (See the deeper treatment in Insight - Your Second Brain Isn’t a Search Engine — It’s an Inference Engine and Insight - The Living Knowledge Base in Action — From Transcript to Intelligence Graph.)
3 — Create Once, Multiply Everywhere
Lou framed the hub as a demonstration of his content philosophy: “If I have a good idea and I create some content once, I want to get as much leverage out of it as possible.” The mastermind transcript becomes recaps, then anything repeatable becomes commands and skills; collections capture his build narratives; and the same source could spin out YouTube scripts, LinkedIn posts, and 60-second clips for Shorts/TikTok/Instagram. The division of labor is the point: let AI do the repetitive administrative/operational work; keep the creative, thought-leadership layer for yourself. For anyone without content yet, his recommended starter workflow: produce content this way, inject schema, publish a sitemap. (Already deeply covered in Insight - Turn Every Conversation Into a Content Engine With AI Synthesis.)
4 — The Offload Pattern: GLM as a Cheap Second Worker
Donald shared a workflow win: GLM 5.2 (via OpenRouter, “almost as good as 4.8” per Nick Saraev, far cheaper). His move — use Claude/Opus to design the work plan, then send the heavy lifting to GLM running in a second terminal while he keeps planning in Opus. “I never got interrupted once. I left GLM on the heavy lifting and worked in Opus planning ahead.” On a big project that would have blown through his Claude credits several times, GLM cost ~$2–3 for the day.
Lou’s counterweight: GLM is 2–3× slower, it’s API-billed (his experience: ~32/day → ~$1,000/month at 8 hrs/day), and OpenRouter has no guaranteed uptime. “Open source is free to download, but you’ve got to host it and get compute on it.” His framing: it’s a great backup and offload — pop over when you hit Claude’s limits, since it’s cheaper than overage credits — but not a daily Claude replacement yet. (See Insight - Opus Plans, Codex Executes — A Cross-Model Division of Labor for Building and Insight - AI Sovereignty — Build Interchangeable-Intelligence Harnesses So No Vendor Owns Your Workflow.)
5 — Build the Tool the Vendor Won’t (Scott’s HPC Story)
Scott’s example was the practical highlight. His organization deprecated an old system that reported HPC job parallelism; the new GUI replacement, built by two or three developers, prioritizes the highest-volume needs — not “Scott’s weird query.” Instead of waiting on a feature request, Scott sat down with Claude, explored a database he had no mental map of (Python doing the SQL: list the tables, count the records, characterize the fields), and walked out with a script that characterized 67 million completed jobs in a 2.5-hour overnight run — answering a question he and his boss had chased for months. Bonus engineering insight from the chat with Claude: run the computation inside the SQL server rather than pulling 30GB of records out.
💡 What This Means for You
Make a list of every “I wish the tool just did X” you’ve stopped noticing. The economic floor for building a bespoke tool has dropped to near zero — your niche need is now a two-hour conversation away from solved.
Deep Dive: Insight - Build the Tool the Vendor Won’t — AI Closes the Long-Tail Gap
6 — Don’t Rent Autonomy — Replicate the Loop (Dirk’s Hermes Story)
Dirk had been trialing Hermes, an autonomous-agent product, and balked at paying twice — Claude plus API cost. So he asked Claude Chat to write instructions making Claude Code “work like Hermes,” dropped them in, and got a run that “wasn’t asking anything anymore — one button, it does everything,” working four hours untouched.
Lou unpacked why: the language model (Opus) is constant; the harness — Claude AI, Cowork, Claude Code — is the system prompt, memory rules, and command vocabulary that shape behavior. Hermes’ autonomy is, at root, a loop plus a memory layer and a “don’t ask permission” posture. Dirk had inadvertently reprogrammed his own harness to behave like the product he almost rented. Lou’s advice: capture that behavior as a skill (or in your global .claude config behind a trigger word like “iterate” or “go until done”) so you can summon it on demand — “and you don’t need Hermes.”
💡 What This Means for You
When an agent product wows you, ask: what behavioral pattern is this, and can I express it as instructions in the harness I already own? Usually you can. Reserve real purchases for proprietary data or compute you can’t replicate.
Deep Dive: Insight - Don’t Buy the Agent, Replicate Its Loop — Extend Your Own Harness
7 — Write Loops, Not Code
Lou generalized Dirk’s moment into a paradigm. Channeling Claude Code developer Boris Cherny — “I don’t write code anymore, I write loops” — the atomic unit of AI work is shifting from the instruction to the goal-seeking loop: figure out what you need, make a plan, create an evaluation rubric for “done,” and keep looping until it’s met. The power compounds because loops nest — a writing team is loops within loops (research until acceptable, drafts until acceptable, headline until acceptable), each forcing more thought at its layer. Two control knobs: a stop condition (a loop without one is a runaway) and effort level — which raises persistence, not intelligence. “Not every problem deserves loops within loops at high effort,” but when output is close-but-not-there, those are the knobs to turn before assuming the model can’t do it.
Deep Dive: Insight - Write Loops, Not Code — Goal-Seeking Loops Are the New Atomic Unit of AI Work · related: Insight - Every Autonomous Loop Needs a Termination Condition — Design It First, Not Last, Insight - Effort Can’t Buy Reasoning — The Model Sets the Ceiling, Effort Sets the Depth
8 — Grill Before You Build (Matt Pocock)
Lou’s resource pick: Matt Pocock — a 15-year coder who teaches with unusual clarity and ships a strong library of skills on GitHub. The one Lou keeps using is “grilling”: a short prompt that has the AI interrogate every aspect of a design exhaustively before you act — blind spots, failure modes, what you actually want — a depth-first / tree-of-thought walk that ends with you both on the same page. Pocock chains it into a full hands-off dev pipeline: grilling → product-requirements doc → GitHub issues → triage → each issue coded in its own branch → PR check → merge to main. “He’s never coding. His work is all up front.”
Lou’s adaptation for knowledge entrepreneurs: keep the grilling, but swap the PRD for a content brief and feed it to your writing team — “what would a skeptic say? what would a first-time reader ask? did you look up that research?” Exhaust the branches, then turn the session into an article brief. This extends the vault’s existing Insight - Explore the Whole Tree Before You Write the PRD with a named, downloadable skill and a content-workflow port.
9 — NotebookLM + Claude: Is the Integration Worth It?
Don raised connecting Claude to NotebookLM via an MCP to drive analysis from Claude. Lou had tried it: “crazy slow” — NotebookLM has no formal API, so the MCP works around it (3-minute retrievals killed it for him). The trade-off he laid out: NotebookLM gives you a real vector database and its slide/report generation, but adding MCP + inference just to retrieve units may not pay off versus dropping files into a Cowork/chat project (also vector-backed) or using NotebookLM directly for what it’s good at. Kasimir uses a NotebookLM connection as one part of his “alpha memory” — a wrap-up skill that summarizes a session and uploads it to a notebook — and offered to share the skill to the repo. (Context: Insight - NotebookLM Transforms When Fed Structured Narrative, Not Raw Content, Insight - MCP vs API vs Skill — Three Patterns for Exposing Your Knowledge to Clients’ AI.)
Community Corner
- A hard week for Lou. His father-in-law’s cancer returned in a big way; the family flew back from Thailand to get care in Toronto. On top of that, his mom’s 90th birthday this weekend and a wedding celebration for his nephew on Sunday. Understandably “a little stressed out” and low on AI-tinkering motivation — yet still showed up with a full hub tour. Bally’s line landed: “that was your quiet week, Lou.”
- Gears cohort. Lou planned to onboard the Gears-package members (repo access + a 30-minute overview) but deferred the group walkthrough since Elizabeth was absent — he’ll do it with Donald, Ri, Elizabeth and others together and share the recording.
- Ambient intelligence / plugins. Lou settled on the plugin mechanism as the way to distribute the ambient-intelligence setup — should be ready next week.
- Fasting banter. Lou forwarded Donald’s book to his wife, who misread “7-day” as 77 days of juice fasting — much laughter and “put Dirk on the fast.”
Links Shared in Chat
- aimmhub.coachlou.com — the unified knowledge hub
- Minimalist theme (work tree preview):
https://worktree-minimalisttheme-aimm-hub-astro.accounts-0c6.workers.dev/commands/tree-of-thought-prep/ - Matt Pocock on GitHub: https://github.com/mattpocock
- GLM 5.2 via OpenRouter (cheap offload model; “almost as good as 4.8” per Nick Saraev)
/grill-me— spotted by Scott in a Nate Herk YouTube video- Scott shared a Mastodon link (nixCraft): https://mastodon.social/@nixCraft/116804588517754145
Try This Before Next Session
- Find your long-tail tool. Pick one “the software won’t do X” itch and turn it into an exploration session with Claude — diagnose the data first, then build.
- Replicate a loop, don’t rent one. If any agent product has tempted you, ask Claude to reproduce its behavior as a skill in your own harness, and test it on a real task.
- Write one task as a loop. Take something you babysit step-by-step and rewrite it as goal + rubric + stop condition; raise effort before adding instructions.
- Grill something. Before your next build or article, run a grilling pass to surface blind spots before you commit.
Open Threads
- Which hub design ships permanently (Claude-style sidebar vs. minimalist) — leaning sidebar.
- Next/previous navigation for leaf entities in the knowledge graph.
- Kasimir’s NotebookLM→Claude “alpha memory” skill — to be shared to the repo.
- Plugin-based ambient-intelligence distribution — targeted for next week.
- Gears cohort group onboarding — to be scheduled once the four/five can gather.
Navigation: index · Previous: 2026-06-18_Mastermind