ceo-assist Eval — 2026-07-20

External eval of this vault/repo as a system. Not a product review for others — a mirror for you.

Verdict: Keep. Simplify and rebalance. Do not rewrite. Do not archive.

Connects to Blindspots — the library-vs-OS imbalance and craft-as-Priority-#4 are already named there. This eval is the structural version of that diagnosis.

What it actually is

Not a multi-agent “CEO assistant” kit.

Layer	Role
Markdown vault	Source of truth (People, References, Learning, Books, SupportWire, Team, Circa, …)
CLAUDE / root `CLAUDE.md`	Operating prompt — co-pilot that makes you sharper, not agreeable
`web/` SvelteKit	Private reader UI → ceo.imkarthikk.com (Access-gated, prerendered CF Pages)
`tools/claude-opener/`	Bridge: site → Ghostty + Claude at vault path
Blindspots	Self-mirror (the conscience file)

Tagline in site config: “The operating manual, the knowledge graph, the mirror.”

Write path: Claude only (no Obsidian-as-editor, post–June 2026 intent).
Read path: the site + graph + Pagefind + “Open in Claude.”
Sister vault: Personal (~/Developer/Personal) = life/home; this = CEO hat (FeatureOS, SupportWire, team, Circa). Personal notes can be rsync’d into the site.

Repo facts at eval time: private · created 2026-03-16 · last push 2026-06-22 · ~648 files · ~600 md · primary code language Svelte.

Architecture (ponytail)

Claude writes markdown → git
         ↓
content.server.ts walks vault (+ personal)
         ↓
static Svelte site → Cloudflare Pages
         ↓
Cloudflare Access (ops, not in-app auth)

Solid, small, solo-shaped. No CI, no tests, no root README — tribal ops.

Strengths

Clear interaction model — Claude writes, site reads. Most vaults die in app limbo; second app deliberately killed.
CLAUDE.md is real — principles, craft bar, “when I’m drifting,” product stories not feature lists.
Blindspots.md is the best artifact — builder-over-seller, refinement-as-procrastination, consumption-as-action, written clarity as shield.
Site is shipped software — wikilinks, backlinks, graph, tags, search, deep-links. Craft matches taste.
Content density is real when it matters — SupportWire context/launch notes, team feedback consolidation, Learning reflections that score moats.

Weaknesses / risks

Issue	Why it matters
Name mismatch	`ceo-assist` sounds like a product; it’s a private OS
Library >> OS	~342 References, ~117 People, ~47 Learning, ~39 Books vs thin SupportWire / Team / Circa — no FeatureOS folder
Doc drift	`CLAUDE.md` still mentions `Resources/`; `Home.md` still has dead Dataview
Security is one gate	Sensitive team feedback + GTM in static HTML; confidentiality = private repo + Access only
Skill outside repo	`/summarize` is the ingest engine but not versioned here
System feeds the blindspot	Graph polish and more Learning folders are the path of least resistance — already called out in Blindspots

June 17 commit burst (graph, nav, YouTube embeds, SupportWire partners, Circa decisions) is peak pattern: shipping and building the cathedral the same day.

Scores

Dimension	Score
Idea	8/10
Implementation	7/10
Docs	5/10
Operability	5/10
Security hygiene	4/10
Uniqueness vs agent kits	8/10

Overall: ~~6.5/10 as a productized system · **~~8**/10 as a personal tool kept in its lane.

Top 3 next moves (shortest path)

Homepage = weekly OS, not library index — thesis, top 3 revenue actions, SupportWire pipeline, open team actions. Graph is not the hero. Fix dead Dataview in Home.md.
Thin FeatureOS/ + feed SupportWire — cap new Learning unless each ends with a done “one thing” tied to those folders; align CLAUDE.md paths to real folders (Learning/ not Resources/).
Security + ops hygiene once — confirm Access on prod and previews; 15-line root README; optionally exclude Team/CEO-Feedback/ from the static walk as defense-in-depth.

What this reveals (from the repo alone)

Taste-first founder who will ship a private knowledge site with graph view before the GTM machine is boring and repeatable.
High metacognition — Blindspots and team feedback show the problem can be named; open question is naming → pipeline.
Bootstrapper running two products (FeatureOS profitable; SupportWire GTM) with distribution as the admitted weak moat.
Risk: excellence applied to the vault can become a high-status form of Priority #4. This repo is both the cure (mirror, launch plans, partner outreach) and the temptation (500+ graph nodes).

Bottom line

Good personal CEO OS with a real private web frontend — not an agent product. Keep the architecture; starve library growth; feed FeatureOS/SupportWire operating notes and weekly action; close doc/security gaps without a rewrite.

The laziest correct strategy is the one Blindspots already recommends: fewer beautiful artifacts, more conversations that move revenue.