I use Claude Code the way some people use a second employee. Not as an assistant — as infrastructure. It runs tasks, manages files, deploys code, posts to LinkedIn. The session is a system.
This is what that looks like on a productive Tuesday.
What we shipped
Free Claude Code Setup Auditor
vibetokens.io/tools/claude-code-audit
Paste your ~/.claude/settings.json. Get a 0–6 power score, a breakdown of exactly what's configured and what isn't, and a specific fix for every gap. Runs in the browser. Nothing sent anywhere.
Six checks:
AUTOCOMPACT_PCT_OVERRIDE— delays context compaction from ~70% to 88%. More working memory in long sessions.MAX_THINKING_TOKENS— the internal reasoning budget. Default is lower than optimal for hard problems.IDLE_THRESHOLD_MINUTES— how long before a session goes stale. Default cuts out too early.skipDangerousModePermissionPrompt— without this + an allow list, every tool call interrupts automation.permissions.allow— what tools run automatically without manual approval.hooks— Stop, PreToolUse, PostToolUse event automation wired into the session lifecycle.
Most setups come back at 1–2/6. The fixes are 10 minutes of copy-paste work.
@vibetokens/mcp — MCP Server
vibetokens.io/mcp | npm: @vibetokens/mcp
One JSON block in your Claude config. Restart. Done.
{
"mcpServers": {
"vibetokens": {
"command": "npx",
"args": ["-y", "@vibetokens/mcp"]
}
}
}
Claude can now:
- Audit your
settings.jsonwith the same engine as the browser tool - Pull the complete reference of undocumented CLI flags, env vars, and 26 hidden slash commands
- Get patterns for CLAUDE.md architecture, coordinator mode, hooks, the
--effortflag - Surface OPERATOR details when your audit shows gaps
The difference between the browser tool and the MCP server: the browser tool is a page you visit. The MCP server lives inside Claude — you stay in your session and just ask. Claude calls the tool automatically when you ask about your setup.
vibetokens/skills — Claude Code Skills Package
github.com/vibetokens/skills
npx skills add vibetokens/skills
Three skills, installs globally, works across Claude Code, Cursor, Codex, and 40+ other agents:
remotion-video — Create programmatic videos with Remotion. The skill teaches Claude the full API: interpolate, spring, frame timing math, composition structure, render commands, and VT brand specs. You prompt. Claude builds the animation. You render to MP4.
vt-content — Blog posts and LinkedIn content in the Vibe Tokens voice. Enforces the voice rules (no ICP language, no generic AI copy, architecture-first framing), the MDX frontmatter format, and the FAQPage schema structure.
vt-blog — Full publish pipeline. Write → generate cover image → commit → push → force-deploy if needed → verify the URL returns 200. One skill, end to end.
The skills package is a public repo. No registry, no submission — anyone installs it with one command. It appears on skills.sh automatically from install telemetry.
Tools hub + nav + footer
Built /tools as the central page for all Claude Code tools. Swapped "The Layer" for "Tools" in the main nav — the series isn't a priority right now, the tools are. Added a Claude Tools section to the footer: auditor, MCP server, all-tools link.
Three blog posts
- We Built an MCP Server That Turns Claude Into Your Claude Code Config Guide — architecture and distribution rationale
- I Installed the MCP Server and Claude Started Answering Differently — first-person experience, written for HN and the Claude Code subreddit
- What Your CLAUDE.md Is Missing — the five-component framework for CLAUDE.md as an operating system
Fixed a persistent Vercel build failure
Every git-push-triggered build had been failing with a TypeScript error:
./app/operator/OperatorClient.tsx:13:9
Property 'className' does not exist on type 'IntrinsicAttributes & Props'.
The CheckoutButton component had className added locally but never committed to git. npx vercel --prod deploys local files so it always worked. Git-triggered builds deployed from git — which had the old type without className — so they failed every time with a 31-33 second timeout.
Fix: commit the file. One line. Working tree and git were out of sync and had been for days. All git-push builds are now passing at ~57 seconds.
What's still pending
npm publish @vibetokens/mcp — the package is built locally and the landing page is live, but the npm package isn't published yet. Anyone trying to install it hits a 404. This needs npm login then npm publish --access public from C:\Users\jason\vibetokens-mcp\.
Cover images for the blog posts — they're using a fallback. The Gemini generator needs to run for each one.
The distribution logic
Three channels, same content surface, three different installation contexts:
- Browser tool — ranks in search, funnels to OPERATOR. Zero-friction: visit, paste, done.
- MCP server — lives inside Claude. Claude surfaces OPERATOR when your audit shows gaps. Runs every session.
- Skills package — lives inside every agent. VT voice and publish pipeline baked into the tool layer.
All three are free. All three reference OPERATOR. The paid product ($147) is the fix for every gap they surface.
The MCP server and skills package are the longer play. Once installed, they're in the session permanently. They don't require a return visit. They compound.
That's one day of building.
