Skip to content
Vault OSGitHub

Claude Code · Codex · Obsidian

Claude Code,
run like a
design studio.

Every task goes to the right model and agent, starts with a small slice of your vault, and is reviewed before you see it. Tuned for product UI/UX design and frontend.

Agent types per model

Task

Build a pricing block

Haiku

agent types

Sonnet

agent types

Opus

agent types

You

questions

Opus writes the spec, then Sonnet builds it, then Haiku checks it, then Sonnet reviews the look, then Opus reviews the whole, then you answer its 4 questions.

Say this in Claude Code

Install https://github.com/mrDesign-ww/vault-os

Open source · MIT · macOS and Linux, Windows through WSL

How it works

One task, start to finish.

Follow one pricing block, for desktop and phone, through every agent that touches it.

HaikuSonnetOpusYou
Step 1 · Main session

You ask in one line.

The main session turns it into a brief: the goal, when it counts as done, the result file, up to three documents to read, and what is out of scope.

Step 2 · Opus · product designer · effort xhigh

product designer · effort xhigh

A designer agent writes the spec.

Every debatable spot gets a number: size, token, radius, state. A skill router loads one to three skills that fit the job, no more.

Step 3 · Sonnet · builder · effort medium

builder · effort medium

A builder agent builds exactly that.

A fresh agent gets the spec and up to three documents, nothing else. No leftover chat from other jobs.

Step 4 · Haiku · checker · effort low

checker · effort low

A checker compares it with the spec.

Fast and cheap: missing parts, empty areas, sizes that miss. When it is unsure, it hands the call up to Sonnet.

Step 5 · Sonnet · look reviewer · effort medium

look reviewer · effort medium

Fix lists go straight back to the builder.

Looks are judged part by part. Two fix rounds at most: if a problem survives both, the next round must use a different method.

Step 6 · Opus · reviewer · effort xhigh

reviewer · effort xhigh

One senior review of the whole.

The assembled block gets one Opus review at 1440, 768, 390 and 320, plus a final check before it reaches you.

Step 7 · You

One card per task, its questions in order.

A task's questions wait in one decision card, in order, each with options and a recommended answer. Your answers are applied and saved to your vault for next time.

Routing

The right model for every job.

Small models do mechanical work. Mid models build. The strongest model designs, plans and reviews. Pick a task to see where it goes.

Sonnet · builder · effort medium

All agent types

Haiku

low
runnerRuns a ready script or test and reportslow
checkerFirst objective check of a render or screenlow
locatorFinds where code lives and returns file:linelow

Sonnet

medium
builderBuilds from a spec or a fix listmedium
workerDeploys, exports, captures and syncsmedium
look reviewerJudges the look of one partmedium

Opus

high, xhigh for design and reviews
product designerUX, specs and sketchesxhigh
architectArchitecture, money and security code, new methodsxhigh
reviewerBatch code review and whole-screen reviewxhigh
general purposeResearch and synthesishigh
  • Unsure between two tiers? Take the higher one.
  • One task, one fresh agent.
  • Every brief names at most three documents.
  • Agents report back in 15 lines or less.

Context

Context is built per task.

In one long session, the rules file, the whole chat and every document you pulled in ride along with each step. Vault OS starts each agent fresh with a small slice of your vault, picked for its role and project.

Most of it is noise for the next task: old plans, research, notes from other projects.

In the prompt

One long rules file
The whole chat so far
Docs you @-mention, in full
Research reports
Old plans
Work and personal notes together

One session

Context

everything, every step

Same project, different slices. The budgets are hard limits: the command line cannot raise them.

Your vault

Design decisions
Design specs
Research
Implementation state
Handoffs
Architecture

Designer agent

Project · Pricing

Up to 8 notes

characters max

Pulls Design decisions, Design specs, Research

Builder agent

Project · Pricing

Up to 8 notes

characters max

Pulls Implementation state, Handoffs, Architecture

Stale index? Retrieval stops: no answer rather than an old one.

Skill descriptions are capped at 300 characters and the whole listing at 5% of context. A router picks one to three skills per job.

Briefs name up to three documents. Research is distilled into short guides that agents read by section.

Orchestration rules load only in the main session. Before Claude Code compacts a long chat to free room, a snapshot of 1,000 to 2,000 tokens saves your last words, the running agents and the state files.

Your vault keeps separate zones, such as work, personal and client. A note from one zone never reaches a task in another.

Quality

Reviewed before you see it.

Cheap checks run on every part. The expensive model looks once, at the whole. And no fix loop runs forever.

Build
Check
Look
Assemble
Review
You
New method

Fix list

Round 1Round 2

Still failing? New method: redraw whole, other tool or route

Build
Check
Look
Assemble
Review
You
New method

Fix list

Round 1Round 2

Still failing? New method: redraw whole, other tool or route

Build
Check
Look
Assemble
Review
You
New method

Fix list

Round 1Round 2

Still failing? New method: redraw whole, other tool or route

  • Parts: Haiku checks what is measurable, Sonnet judges the look. Opus never reviews single parts.
  • Compositions: one Opus review covers everything together, plus a final check before you. Any change nobody asked for counts as a major issue.
  • Code: small changes get automatic checks. One Opus review covers each finished batch.
  • Two rounds per part. After that, the reviewer writes why both failed and the method changes.

Guards

Safe by default.

Hooks sit between the agents and your machine. They stop the few actions that are hard to undo.

If a guard itself fails, the action goes through: a broken guard never blocks your work.

  • Blockedpkill -f nodeStop only processes you started, by their id.
  • Blockedrm -rf build/*No deletes by glob pattern.
  • Pausedweekly usage 98%Work stops, a handoff note is written, and it resumes after the reset.
  • Blockedread screenshot.png · 3200 pxOn macOS, builder agents get images up to 1600 px. Visual reviewers see full size.
  • Keptcontext compactionThe chat is summarized to free room. Your last messages, agents and state files stay.

Design layer

Built for design and frontend.

The same studio rules, tuned for screens: measurable specs, real widths, reviewers who try to break it.

Measurable specs

Every state, every width.

Default, hover, pressed, active, disabled, empty, loading and error, at 1440, 768, 390 and 320. Every debatable spot has a number.

Skill router

One to three skills per job.

74 skills, grouped by job. For each job a router names the one to three that load and the ones to skip, so style rules never collide.

Figma craft gates

Auto layout, library parts, tokens.

Before anything is reported, two read-only gate scripts check layer hygiene and that nothing outside the task changed.

Frontend review

Frontend checked like a release.

Widths are compared with desktop element by element, at 1x and 2x. Motion gets its own review. Nothing is done until a deployed preview passes at 390 and 320.

Memory

Memory with levels of trust.

Folders say where a note lives. Levels say how far an agent may trust it.

  • Agents load policy (L3) first, then canonical pages (L2), then facts (L1). Evidence (L0) only on demand.
  • Memory an agent infers starts as evidence, outside retrieval, until it is verified.
  • Each zone (work, personal, client) has its own index. If an index is stale, retrieval gives no answer rather than an old one.
  • Claude and Codex share memory only through the vault, never by copying each other's internals.
  1. Buttons: 52 pxL

    Policy

    Rules you approved, bound to your name, the date and a hash of the page. An edit voids the approval.

    Buttons: 52 px
  2. L

    Canonical

    The agreed version: project state and workflows that agents build on.

  3. L

    Verified facts

    Checked, with provenance: where each fact came from.

  4. L

    Evidence

    Raw sources, logs and transcripts. Never edited. Read only when needed.

Evolution

Upkeep that waits for your word.

One command runs a hygiene pass over memory and the wiki. A read-only verifier checks the result before anything counts.

Done

Curator

Audits memory and finds pages at the wrong trust level. Makes only safe fixes and proposes the rest.

Hygiene

One zone at a time, under a lock: dead links, unlinked pages, stale claims. Long logs get a summary; nothing is deleted.

Retrieval refresh

Rebuilds the indexes, validates them, then checks them with a live query.

Verifier

Read-only. Checks that zones stay apart, that no data leaves, that it can be undone and that evidence is kept.

Pass · facts (L1) and canonical pages (L2) update

Policy changes wait for you

  • Learning from past sessions (reflection) is off by default. Turning it on needs your word.
  • Each zone (work, personal, client) is processed on its own. Client data stays on your machine.
  • A procedure that keeps repeating becomes a draft skill you approve.
  • It can propose a policy rule. Only you can approve one.

Your seat · iPhone app in development

You only see decisions.

Questions and finished work wait as decision cards in your vault. The Vault OS iPhone app, in development, brings them to your phone. This is where it stands today.

Answer in one tap.

The recommended answer comes first. Accept it, pick another or add a voice note.

  1. Open the question
  2. Read why
  3. Accept
App clip: answering a question in one tap, sample content

Sample content

See progress at a glance.

Every running job shows its step, its round and what it waits on.

  1. All work
  2. One job
  3. Its parts
App clip: following the progress of a job, sample content

Sample content

Point at what to change.

Mark the exact spot on a result. The note reaches the agent at its next step.

  1. Open the result
  2. Mark the spot
  3. Send the fix
App clip: marking a spot on a result, sample content

Sample content

Know it landed.

Sent, picked up on the Mac, applied: each step comes back with the agent's note.

  1. Answer applied
  2. Steer applied
  3. Still waiting
App stills: receipts for an answer and a steer, sample content

Sample content

Why it matters

What changes over plain Claude Code.

  • model tiers
  • agent types
  • bundled skills
  • trust levels
  • fix rounds per part

  • Routing

    Plain

    You decide which model runs each step, or one model runs all of them.

    Vault OS

    The agent type sets the model and effort: Haiku for mechanical work, Sonnet to build, Opus to design and review.

  • Quality

    Plain

    You review whatever comes back.

    Vault OS

    Every part is checked and every whole is reviewed before you see it.

  • Continuity

    Plain

    Memory is one rules file and a notes file, all trusted the same.

    Vault OS

    Specs, decisions and lessons live in your vault, ranked by four trust levels.

  • Attention

    Plain

    You watch the session and answer in the chat.

    Vault OS

    Questions wait as decision cards with a recommended answer. Drafts never reach you.

Install

Install in one sentence.

Open Claude Code and paste this. It reads the install guide, asks a short setup quiz and creates a new private vault.

Install https://github.com/mrDesign-ww/vault-os

  1. Paste the line

    In Claude Code, in any folder.

  2. Answer the quiz

    Trust levels, zones (work, personal, client) and the optional Codex setup.

  3. Open the vault

    Open the new vault in Obsidian and start working.

What it changes

  • Backs up your Claude and Codex settings before it changes anything.
  • Replaces an older system and keeps every personal file. One command rolls it back.
  • Bypass permissions stay off unless you turn them on.

What you get

  • 10 agent types in three model tiers
  • Memory with four trust levels
  • A design layer for UI/UX and frontend
  • 74 bundled skills, guard hooks and evolution

Needs Git, Python 3, Node.js, Claude Code and Obsidian. macOS and Linux; Windows through WSL.

View on GitHubRead the install guide

Your vault stays on your machine. Vault OS itself uploads nothing.