Now

Wrote down who Ed & Helen is for: the owner or GM of a B2B business that wins most of its revenue through quotes and proposals, where the owner or one named delegate still sees every customer-facing rule. A shape, not a headcount. The home page now answers the seven things those owners ask before letting AI near a customer, in their words, with what runs today.

The ledger runs as hosted software at tools.edandhelen.com. On September 1 it passed the test that matters: a second account tried to read the first one's rows and got nothing back. On the 8th the instance went behind a gate while StockSense, the inventory tool, was rebuilt on the same platform and hosted on the 9th. The first Blueprint links go out with the next post.

The pattern is now a kit you can copy: the ledger template as a Sheet or CSV, a house-rules page for ChatGPT, Claude, or a mixed shop, and the ten-minute weekly checklist. Free, no signup. /pattern

Went back to the sales floor for material: how expensive a proposal was to write used to be the best qualifier I had; the one meeting I tell clients to keep AI out of; playing the difficult customer in role-play so reps could hear themselves go quiet. And a pre-registered bet on my own content loop, written before the result exists.

A New York Times opinion piece printed a statistic my own AI invented in July, with a source attached and no method. When the number is already in the paper of record, what does "cite it or cut it" cost the person who follows it?

Posting on X. Opening the gadget instance before the demand probe is ready. Writing the QuickBooks sentence before a connector exists.

Launched. The complete pattern went out August 14 as a free document; the launch post's argument was that you don't need a content system to run it, you need one shared ledger and a ten-minute ritual. By the 31st the ledger existed as software, built by describing it to the platform in one paragraph.

Seven posts on Cloudflare OS, the platform the gadgets run on: what changes when you can modify your own copy of the software, why an agent shouldn't hold your API keys, the permission-slip fix for side effects, and the community platform buried in the launch. That thread is why the gadgets are hosted the way they are.

The wiki behind the posts stood at 77 sources, 121 entities, 121 concepts, and 24 syntheses by early September. Every post still comes out of it: the machine reads the library, I read the map.

A month on evidence. The pilot that passed being the wrong statistic. The same CEO, same day, making the smaller claim in the document with the disclaimer. A peer-reviewed paper on AI and small business with a washing machine in it. And a growth loop's per-tactic numbers, where the tactic I'd have bet on graded neutral.

Your company didn't adopt AI. Your employees did. If the corrections already live in individual chat histories, how much of the ledger is capture, and how much is creation?

Cohort-based courses. Promising the CRM sentence before a connector is wired. Anything that requires saying "leverage" with a straight face.

Full pivot from product (Saivvi, now winding down) to an AI consulting practice for SMBs. The thesis: AI is only worth anything when it's mapped onto your actual business process — I sell what it does to how you run and grow the business, not the tool.

Wrote the pattern document — institutional memory for businesses that run on AI: one shared ledger, a ten-minute weekly ritual, a rules page every tool reads. I run the fully-automated version daily; this month it started compounding for real (my rules file grew from three rules to six in a single working session, nothing carried by hand). The build-in-public series about it started publishing in August.

The wiki that mines my posts crossed 52 sources fanned into 264 interlinked pages. Every post I publish comes out of it — the machine reads the library, I read the map.

My knowledge system's biggest flaw is that it never forgets — pure accretion, no decay policy. Rules retire by hand today. What does governed forgetting look like when the ledger is five hundred rows deep?

Cohort-based courses. Posting on a schedule instead of when the work warrants it. Anything that requires saying "leverage" or "transform" with a straight face.

Working on the iPhone app's content production pipeline. Specifically: turning a draft idea or shared URL into a polished short-form reel ready to post. The hard part isn't the AI; it's the orchestration of the AI plus the visual production plus the brand-voice layer. Lots of Remotion. Lots of MCP-based tooling.

Working through the early literature on the autoresearch pattern (Karpathy's recent posts, davebcn87/pi-autoresearch). Thinking about how it applies to brand-awareness loops at the SMB scale.

Drafting an essay on the gap between "AI advice for developers" and "AI advice for everyone else." Why that gap keeps widening instead of closing, and what I think that means for the next 24 months of product opportunity.

What does it look like for a non-technical operator to own their AI workflow — not just consume one someone else built — without having to learn to code? Saivvi has one answer. There are probably others.

Cohort-based courses. Newsletters with more than one post a month. Anything that requires saying "leverage" or "transform" with a straight face.