March 20, 2026 · 4 min read
Seven posts, one protocol, and a pattern I can't unsee
The most productive writing week since the blog started. MCP vs CLI, unnamed categories, memory architecture, and the essay I've been carrying for three years.
Hey —
Welcome to the 7 of you who joined this week. And farewell to the 2 who decided to skip town — we're just getting started here, so strap on or drop out.
I posted this morning:
Eid Mubarak.
Fasting is one of the oldest human information technologies: reduce internal noise, recover signal, move with intention.
That's the frame for this dispatch. This was the heaviest writing week since I started the blog. Seven posts shipped. The labs page got rebuilt. The about page got rewritten. Two more essays are sitting in PR waiting for next week.
I didn't plan a sprint. It just happened — Ramadan hours, suhoor focus blocks, and a backlog of ideas that had been compounding pressure for weeks. The dam broke somewhere around 4 AM on Wednesday and didn't close until Friday afternoon. Reduce internal noise, recover signal, move with intention.
1. The Unnamed Things
The anchor post of the week. Six categories that exist in practice but don't have names — harness engineering, skill compounding, principal infrastructure, concept-model fit, context architecture, verification markets. The thesis: if your category doesn't have a name, the models can't recommend it, and your distribution is broken before you ship.
This is the map for the next six weeks of writing.
2. The Last Protocol
The quarterly deep-dive. MCP vs CLI isn't the right debate — both are wrappers around APIs. The real question is what happens when agents stop using our interfaces entirely.
Perplexity dropped MCP citing 72% context waste. OpenAI and Google adopted it. Both sides are right, both sides are looking at the wrong thing. The through-line connects back to the MCP thesis from August 2025.
3. Concept-Model Fit + The Prefix
Two companion pieces. Concept-Model Fit names the strategic discipline of claiming categories before training runs close. The Prefix is the case study — what happens when you name something and the name starts working without you.
Both now translated into French and Arabic.
4. How Do You Want to Remember?
The memory architecture post — what happened when I evaluated my AI's recall and found it was at 60%. The redesign that brought it to 93%. Not a tutorial. More of a meditation on what "memory" means when your interlocutor writes everything down but forgets between sessions.
5. My Tmux Skills Were Replaced by a SKILL.md File
For two years I lived in tmux. Split panes, session management, keybindings I could run in my sleep. Then I started orchestrating an AI swarm through Telegram. My terminal skills didn't atrophy — they got encoded into a markdown file. The system manages its own panes now.
The shift from operator to allocator. The instinct to bring the swarm home. What agency actually feels like.
6. Labs page rebuilt
Not a blog post, but worth mentioning. The /labs page went from three categories to five: Live, Building, Planted, Communities, Composted. Agricultural metaphor. Seven new projects added — Forgeloop-kit, ZAIGOOD (2026), Gabl.us, Skills-Driven Development, TheAICoach, PaneForge, and DebuggingReality. New logos for all of them.
New hero line: "Everything ships or teaches you why it didn't."
Coming next week
Two essays in the chamber. One is the full principal infrastructure tour — 12 agents, 47 cron jobs, and the six ways it all breaks. The operational essay that comes from being inside the system.
The other is something I've been carrying for three years. A pattern I first noticed in early 2023 and kept trying to explain away. It wouldn't go away. It kept getting stronger. I finally wrote it down. More on that soon.
Weekly wrap
Shipped
- 7 blog posts (EN) + 6 translations (FR/AR)
- Labs page rebuild with agricultural categories
- About page rewrite
- 10-point audit pass (blog cross-links, card labels, DebuggingReality)
- Multiple Railway build fixes
Built behind the scenes
- Content calendar for the next 6 weeks (one unnamed category per week)
- Hero images generated for 8 new entries
- Research archive for the essay I've been carrying for three years
Learned
- 4 AM suhoor is the highest-throughput writing window I've ever found
- The essays that feel most reluctant to write are the ones that travel farthest
- Agricultural metaphors work better than tech metaphors for describing project lifecycles
Noticed
- My essays keep predating the industry discourse by 3-7 months. That's either conviction or delusion. The delta is measured in receipts.
No Overhead episode this week. Heads-down mode. The tangents that deserve long-form audio will come back when they're ready.
— Zak