Pieter de Hooch, The Bedroom, c. 1658/1660. National Gallery of Art, 1942.9.50. CC0.
Humans at the gates, not in the typing.
Spec and gates, not chat.
Anthropic's AI-native SDLC playbook keeps the stages and kills the pipeline. Originator plus Claude write intent.md → spec.md → engineer accepts plan.md before any file edit → tests → PR reviewed by an agent against REVIEW.md, human code owner still approves → production hooks past deploy. Maintain closes the loop: a control-band breach writes the next intent with no person in the invocation path. Spec and tickets are the audit trail.
One hour in a Codex "PM Chat" produced a build guide plus Soul.md, Agents.md, and an architecture bible. Fresh chats with their own sub-agents executed it. The PM wrote no code. An external chat checked the PM hourly for drift. Token-heavy. Best Codex quality he has seen. Claude refused the same loop. Keep the spec graph. Add a PM chat that never types and a second chat that audits it on a clock.
The human gate on the same family: grill → plan → one-todo implement → human reads all the code, repeated against a checklist. An explainer, a quiz, and a manual QA list. The agent ships drafts. You still read every line before deploy.
Cabinet puts company knowledge, a working AI team, and live apps in one self-hosted folder. Agents are persona plus memory plus cron heartbeats. Outputs land as durable pages, not chat. The laptop can close and the team keeps working. Same "one place is the company" instinct as Team Brain. Different bet: files-on-disk versus a hosted workspace. Interesting, not an automatic swap.
Every agent company should be inbox-by-default. Mint a real two-way address. Forward into it. Do not attach a personal Gmail.
Steal the gates. Do not buy another chat product.
The box keeps working. Waking hours are for review.
If your company is local agents, closing the laptop lid is an outage. Wrap caffeinate so lid-close stays awake. Cheap always-on box.
Codex on one Mac plus Remote is already enough of a personal agent. One box, one harness. The remote is the phone, not a second org chart.
Waking hours scope overnight work. Two-coffee morning review. Coffee #1 is first-pass on what landed. Coffee #2 validates goals before you scope the next night.
ChatGPT and Codex Remote now ship customizable Home Screen widgets. Voice, Chat, Work, Remote as launcher slots. Remote was already how you start a job from the phone. Widgets make it one tap from the first screen you see.
Same bet all week: the box keeps working. You keep the gates.
Designers who build, on real data.
The future belongs to designers who build. Four always-on design agents huddle against a real asset, not a prettier mock. Experiments makes ideas tangible before the roadmap. Motion God wraps the production animation spec in a localhost playground. Figma Bro does production work via MCP against exact spacing and components. Devbot answers how it should actually get built. More versions. Judgment between them. Work continues while you are away.
Shopify, a year after the AI-not-optional memo: about 60% of designers now push code to production. 5,500+ live prototypes on Quick. The mechanism is internal infrastructure, not a Figma plugin. A working proto on real data beats a prettier mock. The only wrong answer is to do nothing.
Apple demos feel magical because every contact already has a photo. Generate a portrait for every contact so first-run looks finished, not like a blank CRM. First-run should feel like a demo with real people, not empty rows.
GTM-as-bots: prospecting, account tracking, inbox triage as specialized bots that do the work around selling, not a chat thinking-partner. Pattern source. Not a product swap.
Stay in attention. Don't outsource the edge.
Getting work done is five skills — which todos matter, name the result, clarify the approach, give them time, stay in attention as drift shows up. Mechanism is a working-memory doc. Dump every open loop. Bold the important-and-stressful items, especially the avoided ones. Protect a block. Do not outsource the edge to an agent until the human is actually there.
Always-on research desks are a labor story. They are not a position.
Try this
Wire one gate and one overnight loop. Leave the edge human.
- Pick one real path. Force intent.md → spec.md → plan.md before any file edit. Human code owner on the PR. No chat-as-spec.
- Scope one overnight job before bed. Two-coffee review in the morning: what landed, then what to scope tonight. Phone is Remote. Home Screen widget if you want the tap.
- For the product's first run: fill the screen before code. Real people, a filled week, one tap that feels done. Working proto on real data beats a prettier mock.
If you still have energy, write the working-memory dump once this weekend. Protect the block. Do not hand the edge away until you are in the chair.
Until next Friday.