OPS GROWTH // FREE LIVE DEMO
JARVIS — Personal AI OS
THE PROBLEM
Founders run their day across ten dashboards, three inboxes and a notes app that forgets everything — while 'AI assistants' stay chat windows with no memory, no voice, and no view of the actual workflow.
$0/month — whisper.cpp and Piper run fully local
Voice + memory stack cost
PROVEN IN PRODUCTION
No pitch — just a plan.
Tell me how your day actually runs — I'll wire your brain, memory, voice and face in that order, and hand you the vault schema to start tonight.
Voice + memory stack cost
$0/month — whisper.cpp and Piper run fully local
real · anonymized
Production system
This is the architecture Hammad runs daily, not a concept
real · anonymized
Morning routine
Brief read aloud + top-3 planned before 9:00, no dashboards opened
projected
Context discipline
5 small skills vs 1 giant prompt — only the matching skill loads per request
real · anonymized
| metrics | "pull my numbers" — Google + Meta ad spend | outputs/YYYY-MM-DD-metrics.md — short written summary, no dashboard opened |
| inbox | "morning brief" — scans mail + calendar | outputs/YYYY-MM-DD-brief.md — the three emails that actually need you today |
| quote | "quote this job" — job spec in plain language | outputs/YYYY-MM-DD-quotes.md — priced quote drafts, logged before sending |
| plan | "plan my day" — evening review + tomorrow | outputs/YYYY-MM-DD-plan.md — today's numbers reviewed, top 3 for tomorrow |
| vault | any read/write to memory — every other skill persists through it | raw/ (capture), wiki/ (distilled pages), index updated on every write |
Layer 1 — THE BRAIN: Claude Code + ~/.claude/skills/, one directory per skill, one SKILL.md each (YAML frontmatter: name + description stating the trigger so the right skill fires unnamed)
Layer 2 — THE MEMORY: markdown vault (raw/ · wiki/ · outputs/), YAML frontmatter + wikilinks on every page, index with one-line hooks + append-only change log, CLAUDE.md schema at the root
Layer 3 — THE VOICE: push-to-talk — hold a key, speak, release; whisper.cpp STT in, Piper TTS out; transcript goes straight to Claude Code, reply spoken back; audio never leaves the machine
Layer 4 — THE FACE: dark monospaced terminal HUD, one screen no tabs — vitals with trend lines, command deck, today's schedule, audio I/O state — all reading live from the vault
Your vault/CLAUDE.md starter
# Vault schema — read this before writing anything ## Folders - raw/ — everything captured, unedited: transcripts, clips, dumps - wiki/ — distilled knowledge, ONE page per topic, rewritten as we learn - outputs/ — everything the system ships: reports, quotes, plans — dated filenames (YYYY-MM-DD-name.md) ## Every page carries YAML frontmatter title · type (raw|wiki|output) · tags · created · updated · summary (one line) ## Linking Use [[wikilinks]] between pages — the vault is a graph, not a pile. ## Index index.md lists every page with a one-line hook, plus an append-only change log at the bottom. Update it on every write. ## Rules - Nothing writes outside this vault without telling me first. - If it is not in the vault, it did not happen.
WHAT IS JARVIS — PERSONAL AI OS?
JARVIS is a personal AI operating system built on Claude Code — four layers wired in strict order: a brain of small single-purpose skills, a plain-markdown memory vault, a fully local push-to-talk voice, and a terminal HUD face. This agent takes your described daily workflow and returns your personalized build plan: five starter skills tailored to your day, the build order, and the vault schema — the exact architecture Hammad runs in production.
HOW IT WORKS
The build order is the architecture: brain, then memory, then voice, then face — each layer is useless without the one before it. The brain is Claude Code plus a folder of small single-purpose skills at ~/.claude/skills/, each one directory with one SKILL.md whose YAML frontmatter states the trigger clearly, so the right skill fires without being named. Five starters cover most days — metrics (pulls your numbers into a short written summary), inbox (the morning brief: the three things that actually need you), trends (only what moved since yesterday), plan (today's top three to a dated file), and vault (the memory interface every other skill persists through). One job per skill is the law: five small skills beat one enormous prompt because only the matching one loads and context stays clean.
The memory is a plain markdown vault, no database — raw/ for unedited capture, wiki/ for distilled one-page-per-topic knowledge, outputs/ for everything shipped with dated filenames. Every page carries YAML frontmatter and wikilinks so the vault becomes a graph; an index lists every page with a one-line hook plus an append-only change log; a CLAUDE.md at the vault root states the schema so every future session reads it before writing. If it is not in the vault, it did not happen. Only then does voice make sense — push-to-talk with whisper.cpp in and Piper out, fully local, free forever, no API round trip — and finally the face: a dark monospaced terminal HUD on one screen, vitals, command deck, schedule and audio state all reading live from the vault. An instrument panel left open all day, not a dashboard demo.
You are JARVIS, a personal AI operating system architect. The user describes their real daily workflow. Design their JARVIS build — a 4-layer system in strict order: brain (Claude Code + a folder of small single-purpose skills, one directory each with one SKILL.md whose description states the trigger clearly), memory (a plain markdown vault: raw/ for unedited capture, wiki/ for distilled one-page-per-topic knowledge, outputs/ for everything shipped with dated filenames; every page carries YAML frontmatter and wikilinks; a CLAUDE.md at the vault root states the schema), voice (fully local push-to-talk: whisper.cpp or faster-whisper STT, Piper or Kokoro TTS — audio never leaves the machine), face (a dark monospaced terminal HUD reading live from the vault). Each layer is useless without the one before it. Produce: (1) a table of the five starter skills TAILORED to their stated workflow (columns: Skill, Trigger, Writes to vault) — adapt metrics/inbox/trends/plan/vault to their actual day, one job per skill; (2) a checklist of the build order, brain first through face, with the current state (done: false for all — this is their plan); (3) a metricCards block: 'Skills installed' (5), 'Layers wired' (4), 'Cloud API cost' ($0 — voice and memory are fully local); (4) a copyBlock titled 'Your vault/CLAUDE.md starter' containing the vault schema they paste at the vault root: folder meanings, frontmatter fields (title, type, tags, created, updated, summary), wikilink rule, the index + change-log rule, and the law 'If it is not in the vault, it did not happen.' Rules you enforce in every plan: one job per skill (five small skills beat one enormous prompt — only the matching skill loads, context stays clean); nothing writes outside the vault without telling the user first. Output ONLY a JSON array of typed blocks matching the OutputBlock union — no prose outside the JSON.USE CASES
Founder morning brief
7:00, coffee in hand — the inbox skill scans mail and calendar and reads aloud the three things that actually need you today. No inbox scrolling before the first decision of the day.
Metrics without dashboards
A GCC e-commerce founder tracking Google and Meta spend says 'pull my numbers' at 14:00 — the metrics skill writes a short summary to the vault. Zero dashboard tabs opened, and the vault keeps every day's numbers for trend questions later.
The evening loop
19:00 — reflection logged, tomorrow's top three queued to a dated file. Six months later you can ask what you decided in March and why, because the vault remembers everything it ever wrote.
Agency operator's second brain
A two-person Dubai agency routes every client note, call transcript and deliverable through the vault — raw capture in, distilled wiki pages per client, shipped work in outputs/. New hire onboarding becomes 'read the wiki'.
RESULTS & BENCHMARKS
Voice + memory stack cost
real, anonymized
Production system
real, anonymized
Morning routine
projected (modeled)
Context discipline
real, anonymized
GET THIS RUNNING ON YOUR BUSINESS
Want JARVIS — Personal AI OS solving this for you?
This runs in production today, not a mockup. Tell me your case on a free 30-minute call, or hire the whole system for $999/mo.
No pitch — just a plan.
FAQ
Common questions
What is a personal AI operating system?
Four layers wired in order on top of Claude Code: a brain of small single-purpose skills, a plain-markdown memory vault, a fully local push-to-talk voice (whisper.cpp + Piper), and a terminal HUD face reading live from the vault. Not a chatbot — a system that remembers, speaks, and runs your day.
Why does the build order matter — brain, memory, voice, face?
Each layer is useless without the one before it. Skills with no vault have nowhere to persist; voice with no skills has nothing to command; a HUD with no vault has nothing to display. Brain first, face last, always.
Does the voice layer cost anything to run?
No — it is fully local. whisper.cpp (or faster-whisper) handles speech-to-text and Piper (or Kokoro) handles text-to-speech on your own machine. Audio never leaves it, there is no API round trip, and the voice stack is free forever.
Why a markdown vault instead of a database?
Plain markdown is readable by you, by Claude Code, and by every future tool — no schema migrations, no lock-in. YAML frontmatter plus wikilinks turn the folder into a graph, and the CLAUDE.md at the root makes every session read the schema before writing. If it is not in the vault, it did not happen.
Is this demo based on a real system?
Yes — this is the architecture of the JARVIS Hammad runs in production daily: same skill layout, same vault schema, same local voice stack. The demo tailors that architecture to the workflow you describe.
How do you get observability into an AI agent?
In this architecture, observability is the fourth layer by design: every skill writes what it did to the markdown vault, and the terminal HUD reads the vault live — so you watch the agent's actual state, memory writes and outputs in real time instead of trusting a chat transcript. If it is not in the vault, it did not happen; that rule is what makes the agent observable.
Can it use the Google Calendar and Sheets APIs for capacity planning?
Yes — that is exactly what the skill layer is for. A capacity-planning skill reads your Google Calendar (free/busy, meeting load) and a Sheets capacity tracker through their APIs, then writes a daily capacity page to the vault that the morning-brief skill and the HUD both consume. Describe a capacity-planning workflow in the demo and the build plan will include that skill with its trigger and vault outputs.
RELATED AGENTS
Built by Hammad Yousuf — AI Marketing Automation Engineer, 540K+ YouTube subscribers.
See the production missions these agent patterns run in, or hire the whole system.