Founder, CEO & Product Executive

My working notes live in an Obsidian vault that 40+ Claude Code routines read from & write to on a schedule.

• Routines cover product intelligence, competitive scanning, meeting prep, content drafting & a nightly journal that aggregates every other routine's output into one entry.
• Every routine grades its own output against a rubric & rewrites its own prompt from the score. What runs today is not what I originally wrote. I did not hand-edit most of the difference.
• One routine exercises a codebase against a library of use cases & an open-gap backlog every day, so drift shows up as a diff instead of a surprise.
• Meetings are captured by webhook, transcribed & then read back against what the vault already knows, so a decision made out loud lands next to the document it contradicts.
• A tracker watches chat channels, merge requests & standup notes to compare the plan against what actually shipped.
• The vault argues with itself. A contradiction pass looks for claims that no longer agree across documents & surfaces them rather than letting the newest file quietly win.
• One routine audits all the others against canonical patterns. It is the only reason 40+ routines have not drifted into 40 different conventions.
• It monitors the open-source projects it was built on, pulls upstream changes & updates its own methodology to use new features. A system that reads its own influences & evolves, rather than a static set of instructions.
• I run it as an operator, not an engineer. Every routine is a markdown file in plain English, so the constraint is knowing what good output looks like, not knowing Python.

The chat widget in the corner & the fit analyzer both run on Chrome's built-in Gemini Nano. No API key, no backend, no inference bill & nothing for me to log.

• The chat widget answers questions about my background from a knowledge base that ships with the page, inside your browser rather than on a server.
• The fit analyzer reads a pasted role description against my public record in four passes & says where the match is real & where the gaps are. The description never leaves the tab.
• Every model reply passes a banned-pattern filter before it renders. A reply that trips it gets replaced instead of shown, which is the boundary a system prompt on its own cannot enforce. 54 unit tests cover that filter.
• Four suggested openers give the conversation a place to start & retire the moment you send a first message.
• If a browser can't run the model, both features stay silent rather than telling you to go install a different one.

NewbornCare is a second Obsidian vault with its own Claude Code routines, stood up for my son from the same bootstrap kit that runs the system above. I built it to answer one question: how much of the methodology was actually about running a company?

• Answer: none of it. The same patterns moved over unchanged. Daily routines, methodology files, run logs, feedback loops & prompts that grade & rewrite themselves.
• A sleep-schedule projection is built from the tracker's own logged data, then recalibrated whenever I reply with what actually happened, so it converges on the real child rather than a textbook.
• A photo routine turns each day into a diary entry in my voice, pulling poster frames out of videos so a video-only day still has something to narrate.
• An activity routine rotates age-appropriate suggestions grounded in learning-science research, indexed by day so no two mornings repeat.
• So the vault system is not a CEO tool that happens to use Obsidian. It is a general-purpose way to run AI-managed knowledge work. The domain turned out to be the only thing that had to change.

The bootstrap kit is a single Claude instruction that builds a complete AI-managed Obsidian vault from zero: folder structure, governance document, the routine template & the order to build things in. NewbornCare was its first real test.

• It does two jobs: a starting point for anyone who wants their own vault, plus a disaster recovery plan if this machine dies tomorrow.
• It starts with a 10-category interview protocol & a post-interview synthesis step, because a vault built without context gets abandoned inside a week. The interview decides which routines to build, which to skip & the governance facts that wire everything together.
• Two complete worked methodology examples you can copy: a crawler (core routine) & a competitive-intel scan (domain routine), each with Section 0, hard bans, pre-flight checklist & run log. Plus prompt skeletons for both core & domain routines, so you know what the scheduled task actually says.
• A Minimal Viable Vault path: start with two routines (crawler + synthesis) on day two & grow to five as the pipeline earns them. A Day 1 Checklist condenses Phase 1 into nine ordered items.
• 20+ domain routine patterns, six contributing repos (four open-source with links, two patterns to build yourself), fourteen subsystems with a scaling table showing when to build each & a troubleshooting section covering the eight failures that hit most new vaults in week one.
• The published version is the pattern, not my instance. It was written clean rather than redacted, because a redaction is only as good as the pattern used to verify it, which is itself one of the failure modes the kit warns about.