Venture

Seeded from another vault

The first commit copied the templates and notes out of the vault of Pantheon, then Brian's local-first agentic OS, and a briefing, a roadmap, a decisions log and a link gate were stood up the same day. The new repository was declared the canonical home of the guidestones; the old copy became a frozen snapshot. A two-way sync between live copies was rejected as the guaranteed path to silent divergence.

The inheritance purged

Seed-era content that did not belong was deleted: 27 operating-agent personas, the old app's internals, superseded meta documents and stale pointers into the other repository. A generic specialist roster and the venture-coaching skill were kept, and two engineering notes moved into the knowledge wiki. Everything removed stays recoverable from the seed commit.

Brian's call Each item was approved one by one.

The foundation build

Ten web-verified, open-source-first tool-stack notes landed, along with four decision frameworks, four operating procedures, a specification for authoring new guidestones and the stage-then-eject pipeline document. A monthly sweep to keep the knowledge current was scheduled.

An airtight scope

The guidestones had been written around the owner's real company and ventures. Every such reference was swapped for a neutral placeholder, line for line, and each file's added and removed line counts were checked to be equal, proving no content had changed. The owner's live venture records were fenced off outside the generic scope.

Brian's call He asked for a generic launchpad with no real venture data in its templates.

The wiki learns three operations

Brian reviewed Andrej Karpathy's "LLM Wiki" write-up and asked whether the vault should work that way. The structure already matched, so only three operations were added: a new source updates every page it touches, answers worth keeping are filed back as pages, and a structural lint joins the monthly sweep. A separate log file and search infrastructure were rejected as premature.

The first freshness sweep, and competitor intelligence

The first monthly sweep checked the most volatile stack claims against the live web six days after seeding and fixed what had moved. Competitor and content intelligence became its own business function: a stack note, brief and self-audit templates, and a recurring sweep procedure, with filled briefs living with each venture rather than in the generic library.

The app stops being a venture

An operations app (a phone media inbox, project management, budgeting and a launch front end) was staged as the launchpad's first occupant, then reframed the same day: it would not eject, because it is the vault's own interface. The no-building rule gained its one exception. Money became work-scoped, with the code committed and raw financial data kept local and out of git.

Brian's call The reframe was his: the app is the project's interface, not a project of its own.

The root becomes the app

All Markdown content, 353 files, moved into one vault folder with its history preserved, so Obsidian opens that folder and the repository root belongs to the app. The operating documents stayed at the root, where the agent briefing must sit to load.

Brian's call He locked the layout during the planning session.

A tower roadmap, and an audit of the templates

Eight scoped guidestones were ordered for building, from SaaS to regulated health, with agencies folded into consulting and franchising kept as smaller templates. The doctrine was set the same day: templates cover every scale, and a venture prunes, never the template. An audit of the template corpus confirmed 36 findings; all were fixed and independently re-verified, including one citation with invented precision that was cut.

Brian's call He approved the tower roadmap and its build order.

Method A multi-agent audit with refute-by-default verifiers; separate agents checked every fix.

A blueprint, then one safe way to write

An approved design spec and a single-file blueprint, 31 stages in five phases, set the shape: fully local, a FastAPI backend working directly on vault files and an installable React app, with a hosted fallback recorded as a plan B and never built. The first stages built the gates and one atomic write path for the app's writes.

A media inbox the phone can share into

A link shared from a phone lands in a queue; the backend downloads it, extracts audio and keyframes and makes a transcript, then stops at "ready". Turning a video into notes is left to an agent by design. The app became an installable share-target web app.

A component library, chosen

Astryx, Meta's open-source design system, became the app's component layer over shadcn/ui and over staying hand-rolled, themed onto the existing tokens, and the first phase's screens were retrofitted to it. A beta library was an accepted risk on an internal tool. It lasted four weeks.

Brian's call Picked from a side-by-side choice.

Money OS

Envelope budgeting over local files: statement import by CSV or OFX with a preview, explicit decisions on duplicates and a review queue. Categories come from ordered rules, first match wins and the rest stays visibly uncategorized; the app never asks AI to guess about money.

Projects, launch, a metrics deck and a CRM

The remaining phases landed in one day, completing all 31 stages with 194 tests green. A metric's latest value is always its last real data point, and empty states never show invented numbers.

Method Every stage ran store, then API, then screen, tests first, one commit per stage. The first phase was built by Codex in its own worktree and merged after review.

A desktop shell

A Tauri window now opens the app, with the backend frozen into the installer and run as a child process that dies with the window, so it cannot be left running. The window opens only after a health check that also compares versions. The installed backend serves the live repository's files, so the desktop app, the dev server and the phone all see one set of data.

No manual reinstalls

The shell gained two modes: a development mode that runs the repository's live backend on every launch, and the frozen production mode with an in-app update button that rebuilds and reinstalls. Tauri's own updater plugin was rejected; a signed release pipeline buys nothing when the code lives on the same disk.

Brian's call No manual reinstalls.

The first birth: Avalon

Avalon, a nutrition tracker, became the first venture through the intake procedure: a scored intake, kill criteria committed in advance with dated gates, a map of which guidestones apply and an approved spec. It was ejected to its own repository the same day, and its staging bay collapsed to the records that stay behind.

The Resource Atlas, specced and seeded

The vault had deep tool-stack notes but no way to see the whole landscape, and the launch pipeline's "pick a stack" step had nothing behind it. The answer was a zoomable map over one Markdown note per tool, with nothing placed by hand: layout is computed from the taxonomy, so the map reflows as it grows. A malformed entry is skipped and listed rather than breaking the map. 271 entries were seeded that day.

Brian's call Two decisions up front: a reference map plus launch picker, and a zoomable canvas rather than a scrolling page or a hand-arranged board.

A thousand tools, researched, not remembered

Per-section research agents took the catalog to 976 entries across 35 sections. The sweep left out tools it found dead or absorbed, and caught licence changes a catalog written from memory would state wrongly. A gate now enforces one home per tool, matched by name and by URL, which caught one analytics tool filed twice. The launch view opens the Atlas in select mode and writes a project's stack file.

Method Cite-or-drop: an entry is written only from a source fetched in that session.

Toward a merge

The plan became to merge the work side into Pantheon one day, so the app adopted Pantheon's shell as a functional copy: a titlebar with note search and view history, a resizable left rail and a browser for the whole vault. The plan was retired eleven days later.

Brian's call The merge direction was his steer.

The app knows whether it is current

The installed app turned out to be nearly three weeks and about 50 commits behind, missing whole features, and nothing could have said so: the version number never moved and the updater never checked its own result. Each of the app's three parts is now fingerprinted from the source it ships, so staleness is measured, and one script updates only what changed and re-reads the install afterwards. A front-end-only change now takes about 20 seconds instead of a five-minute reinstall.

Open into the vault, on shadcn/ui

The app opened to a pairing screen and nothing else, on the one machine that never needed pairing. The desktop app now opens straight into the vault, and pairing moved into Settings for the phone. The same day Astryx was removed in favour of shadcn/ui, which is Radix behaviour with Tailwind styling, copied into the repository rather than installed. The migration exposed three accessibility defects, all fixed.

Brian's call shadcn/ui as the standard across his apps.

The agent gets eyes

A screenshot script now renders every screen in a headless browser, so checking the interface no longer depends on a visible window. The first screenshots found what no test had: the Atlas painted its text colour as its background, so all 976 tool names were invisible. The same day the front end was split so each screen loads on demand, taking the first download from 1.18 MB to 301 kB.

Method Screenshot before calling a screen done, and look at it.

Eight rulings after the split

Pantheon removed its venture, launch and business-map features, two of which had been reading stale copies of Venture's own templates. Eight rulings followed: venture readiness tracking and the business map were built; a renewals list, a hard-coded department list and an in-app template editor were cut as second copies of facts that already lived elsewhere; and an eight-state project board that had never held a single note was retired.

Brian's call He overruled the recommendation to drop readiness tracking: the venture-coaching skill already specified a ledger that nothing produced.

A venture is a folder

Against real data the new Ventures screen came back empty while three ventures sat on disk: they had notes and a business map but no readiness ledger. A venture is now a folder, and a ledger can be attached to one that already exists. The same day the Atlas filled its own backlog of 74 tools that other entries named as alternatives but had no entry yet.

Five defects from an adversarial pass

A review of the two new modules found five real defects, among them a name that could permanently corrupt a ledger and take down the whole list, and two quick clicks that could silently lose a write. Each was fixed in severity order behind a test that failed first.

Everything is a real file

Three settings lived in browser storage, invisible to the owner, to Obsidian and to git, and lost on any new machine. Settings now live in one Markdown file that explains each key to whoever opens it. A file broken by a hand edit says so instead of quietly resetting to defaults.

Brian's call Anything the app knows is a file a person can open.

Venture stands alone

Pantheon became a separate project, and the merge plan was retired. About fifty comments that explained code by where it had been copied from were rewritten to say what the code does.

Brian's call No upstream, no shared shell, no merge target.

Radix everywhere it reaches

Over two days every interactive control was moved onto Radix, or onto a plain native element where Radix has none, with a standing rule against adding a dependency for one control. The move surfaced real bugs, including a picker that saved a blank choice, and deleted two hand-rolled modules. A screenshot caught a broken row layout that 373 passing tests had missed.

Brian's call Anything that could be Radix is Radix, from here on.

The sweep measures before it checks

The monthly sweep now ranks Atlas sections by how long since their entries were verified and deep-checks the stalest. This pass found the drift a link check cannot see: three tools catalogued as open source that are not, a commercial library listed as free, and five projects that had silently moved to new homes.

Eleven new Atlas sections

Infrastructure, security, data, automation, communication, commerce and five more sections arrived with 371 researched entries in one wave, every one verified against a live source. Two sections were cut short when the session ran out and were topped up straight after.

The SaaS tower, and an assembler

The SaaS guidestone was drafted at 13 chapters and three annexes, together with a script that assembles any guidestone from its outline and chapter files and generates the parts that used to drift: front matter, the dated facts table, the contents. Twelve more towers were outlined over two days, each researched at primary sources.

An audit of the guidestones, finished on the second try

Fixes from an audit of the shipped guidestones landed in two passes; the first was cut off when the session limit stopped the fixers mid-run. Chapter headings were kept byte-identical throughout, because the app keys readiness ledgers on them.

SaaS ships through the gate

The SaaS tower passed a two-reviewer ship gate after its fixes landed, with the regulatory dates in its facts table re-checked at source. Marketplace and consulting towers were drafted the same day, the consulting one on Codex, and entered the gate. The corpus stood at five product towers shipped, two in the gate and ten outlined.

Method Two independent reviewers; both must pass before a template counts as shipped.

The tuxedo

The inherited black-and-gold look was replaced with the project's first design system: black, white and one oxblood red, like a black suit, white shirt and red tie, on a Swiss grid. Gold had come to mean nothing, and nearly a thousand full-colour logos made the Atlas the loudest screen in the app. Emphasis now comes from inversion, colour appears only for real data, destructive actions and focus, and navigation collapsed to the left rail. A phone could never open that rail before; it can now.

Brian's call Chosen from three directions shown side by side.

Diana's rules, written first

Before an in-app agent was built, nine rulings set her limits. The app may call a model only through Diana, money is still never categorized by AI, and every write she proposes is held until the owner confirms it. Finance and contact data may reach only a local model, enforced in code rather than by prompt: a prompt is a request, not a boundary.

Brian's call An in-app agent modeled on the one in Alfred, his personal-side project, with sensitive data kept local.

Diana built

Nine stages shipped in a day. Each chat runs on Claude or on a local Ollama model, every write comes back as a card to confirm or dismiss, and paid calls have a per-turn ceiling and a daily spending cap. The local-only fence is enforced at six points, and tests seed marker strings into the sensitive data to prove none of it reaches a cloud request.

What a real local turn found

With 477 backend tests green, the first real conversation on the local model failed in ways no test had: it filed numbers as memory notes and called tools it had not been given. Offering every tool up front fixed the calls but drifted longer replies into another language, so summoning stayed and her instructions gained a worked example. Replies that claim an action never taken are now flagged. The cloud lane says plainly that it is not configured until a key is added.

Ares becomes Venture

The project was renamed so the name says what it is for: getting ventures from idea to birth. Because "venture" already named several things inside the app, every identifier was renamed by hand, class by class, never by global replace, with no aliases left behind. The app took a new white-on-black mark and version 0.2.0; dated records keep the old name.

Brian's call The rename, and that it would follow Diana.

The cutover

The project moved to its new folder, its Python environment was rebuilt, both test suites passed again (477 backend, 108 front end), and the installed app was verified against the repository. References in neighbouring projects were updated to the new name.