AboutProjectsThe OS The SystemEarly workGet in touch
Built with Claude · Claude Corps Fellowship 2025

Chemist by training. Builder by curiosity.

I'm Vanessa — a PhD chemist who taught myself to ship real software with Claude. Apps families use, free tools for teachers, and an operating system that makes Claude govern, verify and improve itself.

Explore the work → See the system I built
0
Projects shipped
0
Live & testable
0
Automation hooks I built
0
Operating-system modules
Vanessa Silvestre Gonzalez
PhD, Green Chemistry · Ogden, Utah · 🇬🇧 🇪🇸 🇫🇷 trilingual+
About me

From the lab bench to shipping software

I spent years as a chemist — a PhD in Green Chemistry, published in mass spectrometry, building machine-learning classification models for cannabis samples long before "AI" was a headline. Some of it was in QC labs under GMP/GLP/FDA rules, where I learnt auditing, traceability and validation fast — the exact discipline my Claude system runs on today. I've also been a teacher and tutor to students all over the world.

Then I started building with AI — and eventually with Claude in particular — and I couldn't stop. With zero formal programming background, I taught myself to scope, ship and improve real products: apps my family actually uses, free tools for teachers and homeschoolers, and the system that runs underneath it all.

What I care about is useful, honest work that helps people — which is why most of what I build is free. My interests run wide: science and research, education, tools for families, accessibility. Right now, as a stay-at-home mum teaching my child as much as I can before I head back to work, learning and family tools are simply where I started — not the limit of what I want to build.

🎓 Free online teaching 🧑‍🏫 Tools for teachers & homeschoolers 🤝 Exchange teaching ❤️ Charity work 🆓 Free-first software
View LinkedIn ↗
Built with Claude

Things I scoped, shipped & keep improving

Every one of these is a real, working project — most you can open and test right now. They're honest first versions: I note what's live and what's next for each, because the work is ongoing.

The centerpiece

COS — the Claude Operating System I built

It isn't an app you can open — it's infrastructure. A system of rules, automatic checks and self-audits I engineered so Claude works like a careful, senior teammate: it plans, verifies, logs and improves itself, every single session.

🎯
Your request
comes in plain language
✍️
Auto-rewrite
prompt-engineered internally
📚
Playbook loads
right rules, on demand
🔨
Build
with guardrails
🔍
Verify
render + measure, prove it
🔄
Log & retro
never forgets, self-improves
0
Knowledge modules
0
Runtime safety hooks
0
Expert "board" advisors
0
Self-audits / month
a real session · replayed
$ vanessa: "build the fellowship website"
✓ hook · response-checklist → rewrote request internally (Persona·Task·Context·Format)
✓ hook · playbook-loader → loaded building-code rules before touching code
✓ scaffolded project · CLAUDE.md · LOG · HANDOFF · TASKS
✓ verify · render + measure across desktop / tablet / mobile
✓ hook · log-staleness → forced a dated log entry, nothing lost
▸ system stays governed, traceable & self-improving.
Self-taught

A system I built — and still supervise

I taught myself all of this and built my Claude operating system on my own — learning quickly, a discipline carried over from regulated QC labs, where I absorbed a whole role's worth of auditing and traceability in only a few months, complemented by a few generous YouTubers whose systems were less complete than where I eventually landed. By the time I took Anthropic's AI Fluency and Claude 101 courses, the work was already familiar — I passed both 10/10. Their 4D framework didn't teach me these habits; it simply gave names to what I was already doing.

Use it→Automate it→Govern it→Auto process-improvement·human input:ReviewFeedbackDecide
🧪 Lab auditing→my system's twice-monthly self-audits
🗂️ Lab traceability→a master index where nothing gets lost
✅ Validation before release→my verify-and-render-before-ship gate

That instinct came from regulated-lab work — picked up fast — and from teaching myself the rest.

⚙️ Automation

Hooks & scheduled audits run themselves — no reminders needed.

🤝 Augmentation

Claude and I build as thinking partners, every session.

🧭 Agency

The COS lets Claude act, verify and self-correct on my behalf.

🎯

Delegation

deciding whether, when & how to engage AI
"Setting goals and deciding whether, when and how to engage with AI." — AI Fluency
On-demand playbook router
The system loads only the rules a task needs, so Claude engages with the right context instead of everything at once.
Mechanism decision guideBEYOND
A rulebook for choosing the right tool for each job — a skill, an agent, a hook or an automation — not just "ask the chatbot".
Cross-tool handoff briefBEYOND
When my free Claude usage ran out, a guardrail let me delegate to a free model (e.g. z.ai) for debugging without it breaking my rules.
Acts autonomouslyBEYOND
It deploys and runs multi-step tasks end-to-end on its own — with a safety guard that forces a private GitHub backup before anything goes live.
Offload to free-running scriptsBEYOND
Whenever possible, work runs as standalone scripts and automations that execute outside the AI's usage — saving cost and stretching every credit further.
✍️

Description

describing goals to get useful behavior
"Effectively describing goals to prompt useful AI behaviors and outputs." — AI Fluency
Auto prompt-engineering hookBEYOND
A dedicated skill turns my everyday, conversational requests into structured, engineered prompts (who / what / context / format) before Claude acts — and when something is vague, it switches to interview mode, asking a couple of clarifying questions first so the brief is right from the start.
Standing rules + power-phrase library
A reusable description system — plain-language conventions, locked decisions — so I never re-explain the same things.
🔍

Discernment

judging whether the output is actually good
"Accurately assessing the usefulness of AI outputs and behaviors." — AI Fluency
Self-preview: render + measureBEYOND
When Claude builds UI, it previews and measures its own work across screen sizes before showing me — so I don't test every little change. Fewer back-and-forths, fewer tokens, better results.
Facts, never assumptionsBEYOND
It double-checks the real sources before speaking — our own files, live documentation for current info, whether a tool/connection is actually available — instead of guessing.
Adversarial critic + a 9-expert boardBEYOND
An internal red-team and multi-perspective panel — each advisor grounded in real methods — review work before it ships.
🛡️

Diligence

taking responsibility for the outcome
"Taking responsibility for what we do with AI and how we do it." — AI Fluency
5 automatic safety hooksBEYOND
Deploy-guard, log-keeper and an auto-retrospective that — at high context — saves my work, then nudges me to free up room. Diligence I can't forget.
A system that improves itselfBEYOND
After each session it scans our conversation for friction and spots new skills, hooks or fixes worth building — then a scheduled audit keeps everything current. Everything is registered, recoverable, and remembered across sessions.
Built-in review mode for human feedbackBEYOND
Every site I build ships with a one-click review mode (this site has it too): I click anywhere to pin a note, then export the whole list for Claude to action — precise human feedback, no guesswork. It's now a standing rule for every website and page my system makes.
The original skills I built — read them yourself
Not downloaded from anyone. These are mine — open the source and see how they work.
improve-system ★
The self-improvement loop. Scans our work for friction and builds the system's next upgrade — this is the one that makes it improve itself.
Read the source ↗
ingest-source
Digests any source (PDF, video, link) into a clean, linked knowledge base — raw → wiki → schema.
Read the source ↗
ask-the-board
Convenes a board of expert advisors — each with a real ingested methodology — for strategic decisions.
Read the source ↗
science-claim-verifier
My scientist's reflex, automated: it checks every health/science claim against peer-reviewed literature before a product ships — and I built it with its own evaluation tests.
Read the source ↗

Anthropic's courses teach you to use Claude well.
I built an operating system that makes Claude govern, verify and improve itself.

None of it runs unsupervised. I stay in the loop on every build — reviewing the work, correcting it, and stopping to re-steer the direction when needed. The system does the heavy lifting; the judgement stays human.

Where I started

Early projects, other AI tools

Before the system existed, I was already building. These came first — made with other AI tools — and they're exactly why the COS exists now: as I revisit each one, I'm rebuilding it better, the right way.

Honest note: a few of these were debugged with other tools — like z.ai — when my free Claude usage ran out. Building on a budget is part of the story, and stretching free tools is a skill of its own.

Credentials

Certified & always learning

A few of my most recent certifications — not the full list. During my years in industry at Cytiva and Teva I earned many more (GMP, quality and lab-systems training), not all of which live on LinkedIn.

🎓

AI Fluency: Framework & Foundations

Anthropic · passed 10/10 ✦

⚡

Claude 101

Anthropic · passed 10/10 ✦

🧪

PhD, Green Chemistry

Loughborough University, UK

🛡️

Safeguarding (Prevent Duty)

UK Government

Let's talk

I'd love to build with the Claude Corps

I'm a chemist who taught herself to build — someone who cares about doing things carefully, keeping them honest, and making most of them free. I'm still early in this, learning fast, and genuinely excited about what comes next. If that resonates, I'd love to talk.

Email me → Connect on LinkedIn ↗