Empty counts as zero
A user receives a plausible-looking number for an incomplete request.
OPEN WORKSHOP / NO AI ACCOUNT REQUIRED / OCT 2026
VIBE CODING.
REALITY CHECK.
Describe the thing. Watch it emerge. Find what breaks. Then do the decidedly unmagical work of making it good.
A free, hands-on introduction to AI-assisted coding: three runnable starters, deliberate bugs, visible tests, local ZIP export. The draft engine is deterministic and illustrative. It does not connect to an AI model.
◆ Browser-only◆ Nothing uploaded◆ Source code you can own
01 / 03BUILD
BREAK
REPAIR
BREAK
REPAIR
THE COMPUTER IS CONFIDENT.
YOUR TESTS DON'T HAVE TO BE.
YOUR TESTS DON'T HAVE TO BE.
THE PREMISE / 01THE FIRST DRAFT
THE FIRST DRAFT
IS A HYPOTHESIS.
“Vibe coding” is a loose label for making software by describing what you want in ordinary language and using AI systems to produce code. That can lower the barrier to trying an idea. It does not remove the need to understand the result, protect users, or maintain what you publish.
In a 2026 Stack Overflow Developer Survey, most respondents reported checking AI output rather than using it unchanged. That is a useful practice—not a verdict on every tool or workflow. Examine the source ↗
SPECIFYINSPECTTESTREVISEHAND OVER
THE MAIN EXPERIENCE / 02
THE WORKBENCH ✳
Pick a familiar use case. Give the specification some shape. Break the illustrative draft on purpose. Then ship a starter that works.
03complete
micro-apps
micro-apps
02 — ITERATE IN PUBLIC
NUMBER DESK
A small calculation can hide an awkward assumption.
ILLUSTRATIVE OUTPUT, NOT LIVE AI
This first draft adds two values, but silently assumes an empty field is zero.
Loading self-contained example code… LIVE SANDBOX PREVIEWIsolated / no network required
Try the empty case; it may reveal more than the happy path.
TESTS THAT DISAGREE WITH THE DEMO
These deterministic checks inspect the teaching model's known behaviour. They are not a browser automation suite or a security certification. Use the sandbox to observe the same flaw.
0 / 3
Tests not run yet
Execute the checks to make hidden assumptions visible.
REPAIR, DON'T JUST RE-PROMPT
The computer quietly treats a blank as a number. The interface looks fine until a person makes a perfectly ordinary mistake.
Require actual values
Validate both inputs before calculating; explain the correction in an accessible message.
Your first draft is still unpatched.
TAKE YOUR WORK OUT OF THE LAB
Export is real: a portable ZIP with index.html, styles.css, app.js, a README and an honest test plan. The files open on a normal computer.
Ready for a first review
The exported project is a starter, not a production security audit. Test in your intended environment.
The repair has not been applied. You can still export, but the known defect will travel with you.
YOUR HANDOFF CARD
Project: Number Desk · Current state: known defect
NO DATA SENT TO US OR TO AN AI PROVIDERDETERMINISTIC SAMPLE ENGINE ·
Open the model: what is actually happening here?
Each case has one intentionally missing edge-case rule. Draft and repair are fixed, authored JavaScript alternatives. Test verdicts are computed locally using explicit booleans: before repair the edge-case test fails; after repair it passes. The other two tests are stable, illustrative checks. The exported ZIP contains actual application code and a manual test plan, not an LLM output or a regression test runner. No user-entered idea is executed as code. The sample preview runs in a sandboxed iframe without same-origin permission.
Limits: this educational model does not evaluate arbitrary prompts, prove code secure, run a real AI agent, or validate every environment. The source viewer displays exactly the files you can export.
“IT RUNS” IS A MOMENT.
“IT STILL WORKS” IS A PRACTICE.
THE FAILURE LIBRARY / 03
THREE BUGS. THREE BIGGER LESSONS.
The workshop intentionally makes small errors legible. Real software breaks in messier ways, and human consequences are rarely contained within a neat demonstration.
00≠∅
CASE 01 · CALCULATORThe ghost of the empty field.
In JavaScript, Number('') becomes zero. If blank means “unfinished,” a plausible result can be worse than an error. Fix: validate before coercing.
FILTER:0 RESULTS…?
CASE 02 · PORTFOLIOSilence is not feedback.
A filter can return no projects and offer no explanation. Fix: visible, accessible empty states that help a person recover.
₹× 0?!
CASE 03 · KIRANA COUNTERInvalid orders look official.
A quantity of zero can slip into an order list when a form only looks valid. Fix: validate quantities and reject incomplete records before updating totals.
A BETTER PROMPT IS NOT A TEST PLAN.
Who maintains it?
Source code is only the beginning. Record the assumptions, where state lives, and how a future maintainer can change things without fear.
Who can use it?
Keyboard operation, focus, readable errors, reduced motion, and small screens belong in the first definition of done.
Who can be harmed?
Client-side validation is a usability layer, not proof of server-side security. Real data, permissions and transactions need independent controls.
What should stay human?
Deciding which problem matters, interpreting evidence, and taking responsibility for consequences cannot be replaced by a convincing demo.
EVIDENCE DESK / 05DON'T TRUST THE VIBE.
DON'T TRUST THE VIBE.
OPEN THE RECEIPTS.
These readings support specific practices, not a universal claim that AI always helps or always hurts. Survey findings describe respondents, while this site's scenarios are authored teaching examples.
What developers do after AI generates code
Stack Overflow's 2026 survey reports how respondents validate AI-generated answers, including local execution and code review. It is respondent data, not a measure of every developer.
Stack Overflow 2026: AI data ↗Why evaluations of vibe coding disagree
A 2026 research review discusses variation by task, measurement and time horizon. This is a preprint and should not be treated as a final consensus.
Michels et al., preprint ↗Threats a working demo does not reveal
The OWASP Top 10 describes prominent application risks, including access control and injection. The tiny exported starters avoid accounts and servers; adding them creates new obligations.
OWASP Top 10:2025 ↗What validation can and cannot establish
MDN distinguishes useful immediate client-side feedback from server-side safeguards. Validation must be designed for actual constraints, not visual appearance.
MDN: Client-side form validation ↗Accessibility can be tested, not assumed
WCAG 2.2 provides testable accessibility criteria. This site aims for AA-minded design but does not claim a formal audit.
W3C: WCAG 2.2 ↗Why one browser is not enough
Cross-browser testing exposes differences in devices, browser engines and user environments. A successful preview cannot prove universal compatibility.
MDN: Cross-browser testing ↗Accessible controls are an engineering skill
A practical course on HTML, keyboard navigation and meaningful equivalent experiences—not just ticking an accessibility checkbox.
web.dev: Learn Accessibility ↗What a direct-upload deployment accepts
Cloudflare's documentation explains dashboard ZIP upload and folder-based command-line deployment, which are different workflows.
Cloudflare Pages: Direct Upload ↗READ THE EXTENDED GUIDE ↗Last editorial review: 26 October 2026 · Links may change over time
EXPLORE RELATED WORLDS / 06ONE EXPERIMENT.
ONE EXPERIMENT.
A BIGGER UNIVERSE.
Connections worth making, not a random link wall. Only destinations whose content could be retrieved during this build are presented as individually checked.
01 / STEWARDSHIP
The Responsible Builder ↗
What changes after you press publish? Responsibility, maintenance and the people inheriting your work.
02 / REVISIONThe Archive of Attempts ↗
How to separate an observed failure from the story you tell about it—and make a real revision.
03 / INTERACTIVE SYSTEMSSYNAPSE / The Idea Lab ↗
Explore transparent, browser-native instruments that connect writing and technology without pretending to run a hidden AI service.
04 / SYSTEMS DESIGNThe World Engine ↗
See another kind of small, inspectable browser experiment: connected worlds and clearly described simulation rules.
05 / LEARNING THREADSThe Sahilverse Atlas ↗
Browse learning and technology experiments; some listed individual destinations still need deployment verification.
06 / ORIGINThe Main Sahilverse ↗
The canonical home for the wider network of writing, technology, work and experiments in public.
A FIELD GUIDE YOU CAN KEEPTHE GOOD PART SHOULD
THE GOOD PART SHOULD
LEAVE THE SCREEN.
Ten pages: project framing, annotated failures, a test matrix, field exercises, maintenance handover and source notes.
Download the Field Guide ↓