Skip to content
For Claude Code + Obsidian

Give Claude Code a brain it keeps between sessions.

FonteOS is how solo founders and small teams make Claude Code actually remember things. Your projects stay separate, your context carries over, and it's all just markdown files in a folder you control.

No npm. No server. No build step.  —  Clone, open, /start.

You're losing hours every week
re-explaining yourself to Claude.

10 minutes per session, three sessions a day — that's 12+ hours a month just on context. And that's before you count the lost decisions and the projects you can't find anymore.

01

"Where was I? What did we decide last time?"

Every session starts cold. That decision from three sessions ago? Gone. The research Claude did yesterday? Also gone. You're paying for the same context over and over.

02

"Claude keeps mixing up my projects."

You ask about your SaaS and Claude pulls in context from your freelance client. Three ventures, one conversation, zero boundaries. Everything turns into soup.

03

"I asked for a landing page and got a design system."

Without rules, Claude expands scope, adds features you didn't ask for, and burns your whole context window on tangents. You wanted a button. It built a component library.

Three steps. No setup marathon.
You're productive in minutes.

Clone it, run /start, the built-in tutorial handles the rest. No config. No weekend lost to setup.

01

Clone the repo

Markdown files and a CLAUDE.md. That's the whole thing. Open it in Obsidian or any text editor. Works on your laptop, works on a server — anywhere Claude Code runs.

git clone → open → done
02

Run /start

First time? A quick tutorial: who you are, what you're building, your priorities. 5 minutes. After that, /start loads everything instantly.

/start → context loaded → work begins
03

Work → /end → pick up exactly where you left off

/end saves decisions, progress, blockers, and next steps. Tomorrow picks up where you stopped. /audit keeps the vault tidy on its own.

/end → state saved → cold start eliminated

Your ADHD bought 12 domains
and shipped zero products.

You sit down to write copy, notice the font looks off, start researching typography, buy a domain for a type foundry, and two hours later you've got a Figma file for a business that will never launch. We all have the graveyard. FonteOS makes Claude call you out mid-spiral.

Output typing

Ships vs. Builds vs. Learns

"Ships" goes in front of a real person. "Builds" is plumbing. "Learns" is research. The system watches your ratio and tells you when you've been building for three weeks without shipping.

  • Ship queue tracked — alerts when empty
  • Build:ship ratio monitored across all issues
  • Weekly ship streak tracked — consecutive weeks with 1+ ship
Always-on

Pattern monitoring

Five patterns that kill momentum, tracked every session. Claude doesn't wait for your monthly review — it flags them while you're working.

  • Scope expansion — "just one more thing" syndrome
  • Prerequisite chains — "I need X before I can do Y"
  • Abstraction escalation — solving the general case instead of the specific one
  • Build-without-shipping — infrastructure that never reaches a user
  • Conditional commitment — "I'll do X once Y is perfect"
Session protocol

Session deliverables

Before you start working, you answer four questions. Sounds simple, but it's the difference between "I'll figure it out as I go" and actually finishing something.

  • Specific deliverable — not "work on project" but "draft the hero section"
  • Type — ships or builds? Sets the accountability frame
  • Done definition — how you'll know it's finished
  • First 15-minute action — smallest possible on-ramp to start
Drift protection

Nudges and interventions

When you start drifting, Claude doesn't just take notes. It actually says something. Sometimes a gentle "hey, we were working on X" — sometimes a harder stop.

  • Spring isolation — one project per session, cross-edits require confirmation
  • Context budget — ~5k tokens max, prevents over-loading
  • Correction loops — if Claude repeats a failed approach twice, it suggests a reset
  • /end comparison — declared deliverable vs. actual output, every session

Start with one file.
Add complexity only when you need it.

No methodology to buy into. No hours of configuration. Day one is one file and one project. Structure grows with the work, not ahead of it.

Two layers, one system

Strategy (vision, priorities, research) lives in the vault. Execution (tasks, status) lives in your task manager. /start and /end keep them in sync.

Vision → Project → Task

core.md holds your north star. Projects (springs) break it into domains. Claude loads only what's relevant to right now.

Kickstart in 5 minutes, grow for years

Day one you have a core.md and one project folder. Run /start and you're working. Weeks later you might have five projects, research files, and custom agents. You didn't plan that upfront — it grew because you needed it.

Easy to maintain

Everything is plain markdown. No databases, no config files, no lock-in. Monthly /audit catches orphans and drift. Weekly maintenance takes minutes.

STRATEGY LAYER — THE VAULT Where you think SOURCE core.md mission SPRINGS Projects ventures STREAMS Workstreams daily work INTEL Research graduated /start /end EXECUTION LAYER — TASK MANAGER Where you do Projects mapped to springs Issues tasks & subtasks Labels agent workflow Linear / Asana

It knows your project
better than you do.

Per-project context, not your whole vault. Tight context window = better output. Claude loads what matters and skips the rest.

Sourcecore.mdmission + priorities
SpringsProject indexesone per venture
StreamsActive workstreamsdaily work
IntelResearch & referencegraduated
AgentsAI role definitionsspecialized
InboxLoose ideastriage weekly
~5k
tokens loaded per session. Claude knows where everything is. Right files, right project, nothing extra. Less noise means better answers.
No vectors, no embeddings, no database. Just a hierarchy of markdown files that Claude navigates automatically. You stop juggling context — it handles that.

Type a command. Or let
the system call it for you.

Commands automate what you'd do every session. Skills handle specialized work. The system knows when to use them — or you just type a slash.

Session commands
/startLoad context, read session state, pull task manager issues, scope to a project, suggest work.
/endSave session state — decisions, work done, blockers, next priorities. Compare deliverable vs. actual output.
/cycleAutomated daily scan. Classify issues, execute approved work, write a daily report.
/auditMonthly vault health check. Broken links, orphaned notes, pattern health. Flags everything, fixes nothing without approval.
/patternsPattern-breaking self-audit. Ship ratio, scope creep, prerequisite chains.
/reviewTwo-stage review. Fast gap analysis, then deep counter-review.
Skills & modes
/perplexityDeep research via Perplexity. Reddit mining, market research, competitive analysis. Cost-gated.
/youtubeAnalyze YouTube videos via Gemini. Key points, summaries, quotes.
/story-editCraft, review, or restructure stories, scripts, and founder content. Voice-aware.
/humanizeRewrite AI-generated text to sound human. Detects 25+ AI patterns.
/noteCreate a new vault note with frontmatter, wikilinks, and correct placement.
/d-*21 design skills — polish, critique, animate, audit, typeset, and more.
Commands ship with the vault (.claude/commands/) and work immediately. Skills live globally (~/.claude/skills/) and need separate installation. Both are markdown files — read them, modify them, or write your own.

Built for integrations.
Runs fine without them.

Works standalone from day one. When you want to add something, the system guides the setup. Each one takes minutes.

Linear
Task management

Two-way sync. Claude reads Todo issues, moves to In Progress on pickup, posts structured comments, proposes status changes. Projects map 1:1 to springs.

✓ MCP ready
Asana
Task management

Same workflow, different tool. First-run setup asks which task manager you use — or none.

✓ MCP ready
Perplexity
Deep research

Reddit mining, market research, competitive analysis. Cost-gated.

✓ MCP ready
Gemini
Document & video

YouTube extraction, PDF analysis, vision tasks.

✓ MCP ready
Google Drive
Documents

Read Sheets and Docs directly into context.

✓ MCP ready
Figma
Design

Read designs, extract components, generate code.

✓ MCP ready
Calendar
Scheduling

Check availability, create events from the terminal.

✓ MCP ready
GitHub
Version control

The vault is a git repo. Strategy under version control.

✓ Native
Your own
Extensible

Any MCP server works. Notion, Slack, Airtable, custom APIs.

✓ MCP compatible

What this isn't
(so you don't waste time).

×

Not a chatbot

It's not a chat interface. It assumes you're in the terminal with Claude Code already running.

×

Not a RAG pipeline

No vectors, no embeddings, no similarity search. It's markdown files loaded in a specific order. That's it.

×

Not a SaaS

No server, no account, no subscription. It lives in a folder on your machine. You own it entirely.

×

Not opinionated

Use any task manager, any methodology, any cadence. The system adapts to how you work.

Clone it. Start it.
The tutorial does the rest.

First /start detects you're new and runs the tutorial. 5 minutes. Every session after loads instantly.

zsh
$git clone https://github.com/luke-toledo/fonteOS.git my-vault
Cloning into 'my-vault'...
$cd my-vault
$claude
Claude Code ready.
>/start
First run detected. Let's set up your vault.
What's your name? What are you building?