The room where the machine lives.
Door 05 shows the effect on the week; this room shows the shape underneath. Eighteen automations, arranged by where they live in the production cycle and by how soon they can ship. Depth on tap; nothing hidden.
Audience honesty. This door started life unlisted, and earned its spot on the front of the room — the progress inside was too concrete to keep behind a whisper. It's here for whoever wants to kick the tyres — the tyre-kicker's likely Sophie, doing the plumbing for the new site while the rest of the room stays feelings-first. Ron is welcome, of course. So is future-Nick, who'll be reading this to remember what he decided.
Three beats. The machine tends the middle one.
All knowledge work runs the same loop — gather, decide, learn. The engine takes the routine parts of the middle and keeps the routine off Ron's desk. Recording, judgement, taste — all of it stays Ron's.
Guest research, archive cross-refs, a one-page brief in the inbox 48 hours before Ron sits down. Ron reads, annotates, arrives at depth.
Transcript, show notes, blog draft, captions, thumbnail brief, clips, nurture — a single production document that Ron reviews once. Approve, edit, redirect. Nothing sends without sign-off.
Ron's edits become the calibration data — the machine gets a little more Ron with every episode. The archive deepens; every future dossier stands on more ground.
One sentence of context, then no more trend talk on this page. Anthropic's own numbers, June/July 2026: more than 90% of Cowork activity is non-coding, and business operations plus content creation account for roughly half of all usage. Ron's use case is the tool's mainstream, not an experiment.
One recording in. The rest arrives on its own rhythm.
Read left to right. Phases are temporal — the sequence a single episode moves through. Each cell is coloured by tier: what ships now, what waits on a dependency Ron controls, what lives on the horizon. The empty cell in the middle is deliberate — the recording itself stays Ron's.
Automating presence is a failure mode. Riverside is an upstream input, not an automation target.
What changed since April. Nine of the eighteen now ship in the first six weeks — 04 (Guest Curator) and 06 (Voice Model) moved from T2 into T1 because Cowork scheduled tasks shipped on the 7th of July, and because a working preference-compounding pattern went public a fortnight before that. Three more (05, 07, 11) become T1 the moment Sophie's GHL migration lands. Details on every card.
Each one, at the depth its tier warrants.
Click a card to expand. T1 entries are specified at working depth. T2 specifies the bridges T1 outputs rely on. T3 is sketched — enough to see the shape, not enough to over-engineer what lives six months out.
Nine automations. These define whether the engagement lands.
Seven from the April tier, plus two the July capability shift promoted in.
01Research Dossier
02Production PackageTrust △
03Distribution Package
04Guest CuratorTrust △
06Voice Model Refinement Loop
08Thumbnail Brief Generator
09Transcript Pipeline
10Reels & Shorts Generator
12Coaching Brief (pre-call)
Three that ship the moment Sophie's GHL migration lands.
GoHighLevel released an official Model Context Protocol server in June — free, five-minute token paste, roughly forty tools growing to several hundred. The whole reason these three were T2 in April was "GHL integration = architectural build." That build is now plumbing.
05Calendly → Brief, full pipeline
07Nurture SequencerTrust △
11Membership Content Router
One left in this tier after the July shift.
Everything else in the original T2 either moved up (04, 06) or became conditional-T1 (05, 07, 11). This one stays where it was because it earns its complexity honestly.
13Archive Ingestion & Tagging
Five, sketched — not proposed for building soon.
Their purpose on this page is to show how T1 outputs compound into T3 inputs — not to over-engineer what lives six months out. Capacity restoration first; ambition later.
14Corporate Speaking Prospect DB
15Second-Book Research Extraction
16Healthy Bite Repurposing Engine
17Archive Knowledge Base
18Speaker Prospect Enrichment
Thursday. Forty-five minutes. A stack of cards.
The week's promise (door 05) was "one review session, one coffee. Approve — it goes." This is the shape underneath it. Every item that would carry Ron's name arrives as a card with a proposed action. Ron talks his way through the pile — approve, edit, redirect, drop — and nothing sends without his sign-off.
The card-sweep, in a sentence: every artefact that leaves this room passes through Ron's eye first.
Some cards are quick — a thumbnail brief, a clip shortlist. Some deserve real thought — a nurture email to a corporate prospect, a guest outreach note in Ron's name. The three checkpoints below are not procedural. They're the load-bearing walls. Remove any one and the trust structure collapses. So they don't get removed.
Guest selection— 04
The podcast is Ron's voice, curated by Ron. If the machine recommends and automates outreach, the podcast gradually becomes an artefact curated by the machine. That's the wrong direction. What makes this show work at hundreds of episodes is that every guest is a guest Ron chose.
- Bulk approve is not offered — every candidate, separate click
- No outreach in Ron's name without per-candidate sign-off
- Draft outreach is advisory; Ron can replace it
- Persistent "do not recommend" list Ron maintains
- Volume calibrated to Ron's attention, not the machine's throughput
Pre-publish show notes— 02
This is the voice surface. If the package publishes without Ron's review, every downstream distribution piece carries content Ron did not sanction. Voice drift becomes voice replacement — the slowest and most damaging failure in the pipeline.
- 30–45 minutes per episode; design target, not nice-to-have
- No auto-publish at this stage, ever, until at least a dozen calibration episodes and Ron opts in
- Drift surfaced in plain language, not a silent metric
- Voice model version-controlled; Ron can roll back
- Ron maintains a "voice locked" phrase list that cannot be retrained away
Pre-send nurture— 07
Email lands in inboxes with Ron's name on it. A tone-wrong email to a coaching prospect degrades the relationship; to a corporate prospect, costs the pipeline; to a subscriber, triggers an unsubscribe. Unlike a bad draft (which Ron reviews), an automated send lands in people's lives without his eyes on it.
- Pre-send approval by default
- Segment-level auto-send after four clean runs — not campaign-level
- Corporate track stays manual-approve for the first ten sends
- Frequency-cap logic per subscriber prevents campaign stacking
- Diff-log every send — Ron can see what changed even when auto-approved
All three checkpoints sit at the boundary where Ron's voice reaches the world. Guest selection reaches his audience through who he platforms. Show notes reach them through what he appears to have said. Nurture reaches them through what he appears to have written. Protecting those surfaces is not a feature — it's the engagement.
The compounding, traced in four chains.
Static tier separation misses the thesis. The work compounds — each T1 artefact deepens what a T2 automation draws from, which deepens the corpus a T3 automation serves. Four chains trace this explicitly. Two are running-shape today; two are horizon.
What this chain doesn't promise. A book pipeline that writes chapters for Ron. The course-to-book chain Ron already runs in Claude — that's what "already" means; this machine feeds it, and its own e-book-with-embedded-clips shape is his idea, sitting here waiting for a build slot, not shipping in the first six weeks.
AEO — answering-engine optimisation. Ron's marketing friend's thesis, in Ron's own words: everyone is still doing Google organic, but ChatGPT is where the attention is going. The corpus that 09 and 13 build is what makes Ron's answers findable there. Feeds Chain D at the horizon, not the first six weeks.
Amy is the spine of this room, not a target of it. The workflow document she keeps — the one Ron said she updated for the discovery call, at his request — is the artefact every automation on this page is scored against. If the machine hasn't lifted a piece of work she carries today, and hasn't lifted it in a way she can see and audit, it isn't shipping. Nothing here displaces her work; the whole point is to move the drudgery off her hands and leave the judgment on them.
Two, sitting in plain view.
The first has been true since the mockup. The second showed up on July 7th, with a Cowork release. Both stay named — that's the register the whole engagement runs in.
Ron wants less management and more control at the same time.
Those two pull against each other in every non-trivial automation. Visible in the per-episode review, in guest curation, in every nurture send. The architecture doesn't resolve the tension by compromising between the poles. It resolves it by allocating — category by category, artefact by artefact.
Fewer decisions per week. Fewer logins. Fewer approvals. Fewer edge-case escalations. The hours of weekly plumbing collapsed into a shape that fits in Ron's pocket.
Ron's voice stays Ron's voice. The podcast he puts out is the podcast he wants to put out. The corporate pitches he signs off on are ones he would have written. Nothing goes out that he doesn't recognise.
The test. If Ron is asked to review a thumbnail three weeks in, that's a failure. If Ron is asked to approve a nurture email three years in, that's the architecture working.
Scheduled runs store state where Ron's data isn't stored today. [ASK NICK]
Cowork's July 7 release made scheduled tasks and remote sessions real — the reason 04, and the whole rest of the T1 layer, can honestly claim to run without a window open. The catch: those remote runs keep session state in the Claude account, not on the Mac Nick works from. The service-bureau posture — the ongoing fee buys adaptation on Nick's hardware — bends at that seam.
Recurring triggers Ron never has to think about. "It just arrives" becomes literal.
Files-first for outputs (Ron's data stays Ron's files), but the scheduler's own state lives in the Claude account. Options: keep scheduled runs to non-sensitive triggers, or run recurring jobs from Nick's always-on Mac. Nick's call, flagged here so it isn't silently made.
Every automation reads and writes files. That's the whole thing.
No database. No black box. Markdown, CSV, JSON — the shapes any human can open. Amy can audit any decision by reading a file. Sophie can inspect any pipeline by opening the folder. Ron's data stays Ron's files, always.
Why this matters more than it sounds like it does. It's the exit guarantee. If this engagement ends tomorrow, Ron keeps everything the machine built — because everything the machine built is already sitting in his Drive, in the same folder structure his workflow doc points at. The IP is inspectable, the adaptation is the service. No lock-in, no ransom, nothing that behaves like a black box.
Every pipeline step that touches text writes its output to a file first, then a view renders it. The dashboard is optional; the files aren't.
Where the machine is guessing — flagged, not silently resolved.
Every flag below is an invitation to say "no, actually." That's how the engagement calibrates. Ron's edits are the calibration data; his silences are calibration data too, but a spoken correction is a shortcut.
The canonical category count. The room has been saying 21 or 27 depending where you look. Ron used 27 on the discovery call; the catalogue supports both 26 (split) and 27 (with one unpopulated). The flag stays open here because the answer belongs to Ron, not to us.
Where the Holistic Alternative Broadcasters conversation should sit — in or out of this room. Not discussed on the call. Open by design.
Fluoridation as a stance — in the house-style rulebook, out of it, or neither. Same answer: Ron's, not ours.
The path-to-yes on new engagements — a standard we could publish, or the shape Ron keeps in his head. Both are legitimate answers; the flag stays open until Ron picks.
Where scheduled-run state lives. Named in the second tension above. Nick's call before any scheduled task carrying sensitive triggers ships.
The GHL migration timing. Three T1-conditional automations (05, 07, 11) wait on it. What order does she want to build in, and where does this room's plumbing help vs. get in the way?
"The interventions are remarkably simple, sustainable, achievable and effective. And effective, that's really important."— Ron, on the discovery call. The design principle applied to his work, applied here to his week.
Next · 04The working note — the day-one document →