Last time I said there'd be more on the architect/builder side. Here it is: https://youtu.be/rRRsZmKoAho The habit that carried me before AI was building small. Find the smallest piece that feeds the next one, get it working, test it, then add on. Sometimes that meant hard-coding a few test cases just to prove a piece out. It's a big part of why a lot of my work never came back for rewrites. Architect/builder is that same habit with two sessions. The architect session plans with me, asks the questions I'd forget, and never writes code. The plan goes into a handoff file. A fresh builder session reads only that file, tells me how it would build it, and stops until we've looked at it. Then it builds a piece, tests it, reports back, and we decide whether it gets committed. The fresh session matters for the same reason ICM loads one stage at a time. In one 2025 study (NoLiMa), 11 of 13 models were below half their short-document accuracy by 32K tokens. A long planning chat is the same kind of pile. This is where a lot of people get stuck. They know the chat is too long, but they won't close it, because that's where everything lives. So getting organized comes first. Decisions and current state go in files in the repo, and then closing a chat costs nothing because the next session picks up right where the last one stopped. I've reorganized my own setup two, almost three times to get there. If you've been at this a while, or you're heading toward something bigger, it's worth the step back. I also expect this one to stick around. Loop features and slash commands change every few weeks. This is closer to how software shops already run work, and it sits fine next to ICM because the plan lives in a file. Question for the room: the full cycle is a lot for a small change, and I don't always run the whole thing either. Where do you scale it down? Which part do you drop first? P.S. Show your work bit: same deck setup as last time. 31 clicks, 31 slide states, cut straight off the Screen Studio click log. The transcript pass also caught me saying accuracy falls "about 50% of the way through the context window", which isn't what the chart shows, so that line got cut.

I came across a pretty clickbaity video on Claude and file size: Simon Scrapes, "Everything You Know About Skills IS OUTDATED". The headline claim: when Claude opens a long file it runs head -100, so anything past line 100 basically doesn't exist. These ones catch my eye because I'm building these folder setups for actual clients. I use mine every day, but I also want to be handing over the best product I can, so I wanted to know if this was happening to me before I touched anything. I pulled the transcript, then read the Anthropic best-practices page he's quoting. The page is narrower than the video. It says Claude may preview with head -100 when it reaches a file through another referenced file (SKILL.md points to advanced.md, advanced.md points to details.md). Their fix is to keep references one level deep and put a contents list at the top of anything over 100 lines, so a partial read still shows what's further down. The page also has no date on it, so "the rules completely changed" is hard to check. Then I had Claude go through my own Claude Code session logs, 547 of them, and find every time an agent read one of my instruction files partway. What came back: - My skills are pointers. A stub in ~/.claude/skills says "go read the real one in my workspace." That's an extra hop, the exact shape the doc warns about. The agent followed it through 224 times out of 224. - My wiki index is 497 lines. Agents read it whole twice in about 230 visits. Every other time they grepped it or pulled one section, and never missed what they needed. - The actual miss was my writing style guide. 104 lines, one hop from the router. Six sessions read it to somewhere between line 60 and 80 and stopped. The last section, the one that says a rewrite can't change the meaning, got skipped all six times. - One of my video skills, 302 lines, got read in two chunks once. The second chunk stopped right before the gotchas. - And one that was just a bug. My vault's startup instructions said "read the last 10 log entries." The log is newest at the top. 21 sessions did exactly what they were told, ran tail, and got entries from May and June. 10 of them never saw anything current. That one was a one-word fix.

I've been building my second brain for a a couple of weeks now. I'm trying to use it to help me be more organized in my business but I feel like I'm getting more distracted. Does anyone have some tips on this situation? I'm sure it's happened before to somebody else.

I've spent over a decade building software for clients. Small project teams out at the client, with a big core product team behind us. A lot of you are building without ever having worked inside a software shop, so here's the habit from that decade I'd hand you first: https://youtu.be/CnRM0NccyXs Waiting for perfect requirements is how a project stalls out in analysis paralysis. So we didn't. We got the gist, built something rough, put it in front of the client, and then shut up and watched them use it. One client told us flat out they didn't want the imaging side of our system, the part that stores the scanned documents. Theirs was fine. So I built a rough demo: a form with its scanned document sitting right on it, no jumping into their imaging system to go find it. As soon as they saw it, the whole contract shifted to include it. No requirements meeting was ever going to get that out of them. People can't tell you what they want until they see it. That loop used to cost a team weeks per round. With AI it's an afternoon. Most people in the AI space use that speed to finish faster. I'd rather use it to be wrong faster, while being wrong is still cheap and everything is still running on fake data. The video also walks the part of the map most new builders skip after the loops: putting the pieces together, a full run from start to finish, real data with security on, rehearsing go-live. Going from loop one straight to production makes you the TIP guy: test in production. Don't be the TIP guy. Question for the room: who's your first user? If you're solo and building for yourself, who do you hand the rough version to, and how do you keep yourself quiet while they poke at it? P.S. Show your work bit: the slides were an HTML deck with click-through reveals, and the cut came straight off the Screen Studio click log. 27 clicks, 28 slide states, no timeline scrubbing. More soon on the architect/builder side of this, which is the same fresh-eyes idea run with two agents.

I think I have 30-40 hours left of Fable access. My usage has just ticked over to 56% used. I had enough of a build queue that I decided to upgrade my plan just for this month and the Fable-in-subscription window and honestly, it's already been worth it. How those tokens have been spent: - 5% used to complete a system recarve and hardening pass - 35% used to orchestrate a full substrate port (from NotePlan to Notion) *and* a full branch that is future compatible with that state ready to switch to on Friday assuming this week on Notion goes as planned (oh and a full school comms rewrite to go with my brand new gmail set up exclusively for my system to access). I didn't have to build a thing in Notion - Sonnet dispatches did the lot. - 5% used to build the work order stage of my construction crew - 11% used to build the orchestration, build oversight and dispatch stage of my construction crew (and a massive shout out to @Ari Evergreen - I forked Pushing Dispatch_ and that helped me keep this to 11% of Fable orchestrating, otherwise it would have been more) Every stage of the construction crew has been pushed through a live run of the step before it in the chain to produce the document that guided the build. For the work order stage, that meant a PRD from the first stage and a build spec produced by the second stage. For this orchestration, it meant that I had fully sliced work orders, mapped to my schema, clustered to work packets that made sense. 8 stages; that were able to run in an orchestration-style mode (on Fable, I would not have let Opus go like that to be honest) and I came back to high quality work, measured to my own rules and standards and only a few small tweaks that didn't require substantial rework before committing. Honestly, the fact that something of this scale could be overseen with only 11% usage feels like a real testament to the whole chain here. I mean... this is really something. That means I still have 44% of my Fable tokens to go and only 30-40 hours to spend it in... and now we see how much of it I can use before I lose it... Wish me luck.