Different machines. One direction.

Technical deep dive · AI Marketing Agents

Narrow jobs, shared memory.

Every agent is built in the same order, from the bottom up. First the scripts that do the mechanical work, then the skills where the judgment calls happen, and only at the end the agent file that routes between them. It feels slow at first. It pays off the first time a platform changes and only one script needs fixing.

Pipeline

1. 01 Brief A topic, the channels and the brand rules. For partnerships, a short description of who we're looking for.
2. 02 Evidence Studies, public conversations or blind competitor searches. Every fact keeps a link back to where it came from.
3. 03 Produce Several models and tools per asset: Claude writes, image and video models generate, code assembles to brand rules, then a proof-read pass.
4. 04 Engage Social channels, influencers and affiliates run through agents within set limits, with pacing, hard stops and a full action log.
5. 05 Learn Logs, playbooks and the rule base are append-only, so each run starts with whatever the last one learned.

The agents

• Research to Ideas Content strategy Three sub-agents in a row. One finds recent studies, one turns them into 20 or so ideas, and one fact-checks every claim before the list goes anywhere.
• Content Creator Creative production Turns ideas into carousels, single images, infographics and decks, combining language, image and video models. Around 1,200 creatives so far, each through a proof-read.
• Social Media Manager Community Runs the brand's social presence in my voice, scoring what to engage with against my views. Every comment is checked for the usual AI tells first.
• Influencers and Affiliates Partnerships Finds creators and partners who fit a brief, verifies each through four gates, and manages the pipeline from a tracked manifest.
• Competitor Landscape Market intelligence Five research streams that can't see each other's results, so they can't all copy the same mistake. One run mapped 554 companies and 154 real direct competitors.
• Social Listening Audience research Reads public conversations at scale to find what people actually worry about. One study went through about 345,000 posts and comments to find around 3,400 that mattered.
• PR Kit Earned media Drafts press releases as designed PDFs, builds a journalist list with a confidence note on every address, and writes pitch templates.
• Video Ads Paid social creative Makes short vertical ads, from AI-generated presenters to product demos assembled from automatically captured app screens.
• Reporting Data engineering Pulls raw exports from ad, affiliate and analytics platforms, joins them only on declared keys, and computes every metric from fixed definitions.
• Performance Analytics Analysis and recommendations Flags anomalies and outliers, applies a layered rule base from universal to client level, and ranks actions by expected impact. Rules I confirm join the base.

Technical deep dive · VC Catalog

One connected map of the venture market.

VC Catalog runs on two layers. A data pipeline collects firms, funds, companies and people from public sources, resolves them into single entities and links them into a graph of investments and team affiliations. A web platform on top turns that graph into search, profiles, market reports and Connect, which finds the shortest path between two people or organizations. Every change to the data ships as a versioned, checked release.

Pipeline

1. 01 Discover Firms come from public databases and SEC Form D and Form ADV filings (US securities and adviser registrations). A Form D fund size is read as the amount sold, not the target.
2. 02 Crawl Firm and startup websites are read for portfolios, teams and theses. The crawler looks for a site's own structured data first and never bypasses a bot check.
3. 03 Link and dedupe Records join only on hard identifiers, like a domain or an SEC filer number. A matching name is never enough, and every rejected merge is remembered.
4. 04 Classify Claude agents label sectors and business models offline. A local classifier handles volume where it reaches about 90% precision. Every label keeps its provenance.
5. 05 Release Each release must pass 43 health checks. An import that sees an unexpected schema refuses to run, and no sync may delete more than 5% of a table.
6. 06 Serve The release loads into PostgreSQL, the search and graph indexes rebuild, and the site picks it up.

Technical deep dive · AI Software Factory

Gates, not committees.

Models write the code. Everything else is decided by a deterministic controller: which model runs, how much it may spend, when a task gives up, and what is allowed to merge. An earlier version ran for eight days and about 700 agent runs before I stopped it, so nothing here retries forever. Every loop has a counter and an exit.

Pipeline

1. 01 Intake The controller checks the Studio package first. Every file hash must match the manifest and every artifact must be approved. Then it sets up the repo and tasks, no model involved.
2. 02 Plan Four planning passes on the strongest model, each followed by a code check for coverage and cycles. The first milestone is always a walking skeleton: the thinnest version that works end to end.
3. 03 Build legs Tasks run as headless Claude Code sessions, capped in turns and budget. A router hands each one a short-lived key for a single model.
4. 04 Integrate One candidate per project at a time: tests on the merge tree, a test canary, a secret scan, a review if the task is critical. Then the controller merges.
5. 05 Milestone preview Build once, deploy a preview, replay every journey so far on three devices with accessibility checks, then tell me what I can try.
6. 06 Release Scanners, red team, judge. The exact tested image deploys without traffic, is smoke-tested and approved, then goes live, with automatic rollback.

The agents

• Product Design Studio Idea to product package Turns a one-line idea into requirements, UX, UI, architecture, security and runnable acceptance tests, stage by stage.
• Planner Architecture and task graph Runs four planning passes and then reviews its own plan, on the strongest reasoning model available.
• Developers Pool of 2 to 6 Cost-efficient coding models, chosen per run and never swapped halfway. How many run at once depends on how much memory the machine has free.
• Diagnostician Repair after repeat failure Steps in when a task fails the same way twice. It gets one repair attempt. After that the task stops and waits for me.
• Reviewer Specialist by exception Only looks at tasks marked as critical. Reviewing everything is how the committee problem creeps back in.
• Slice Verifier Milestone demo check Clicks through the demo steps on every preview, roughly the way a new user would.
• Red Team and Judge Release gate One tries to break the release candidate. The other, separately, decides go or no-go.
• Health Monitor Controller, not a model Watches 16 warning signs and can pause the whole factory by itself. It can't unpause. That part is mine.