SyntaxPilot
Courses/Vibe Coding/Lesson 40
Video coming soonThis lesson’s video is being made

Start with the notes below; the video appears here as soon as it is on YouTube.

Vibe Coding · Lesson 40 of 40

Vibe coding for jobs and freelancing

portfolio of shipped projectsexplaining AI-built code in interviewswhat employers expectfreelancing gigspricingthe skills to learn next (DSA, system design, a language course on this channel)

Vibe Coding: build real apps with AI · Episode 40 (the last episode of the course). Every agent run below is real: Claude Code 2.1.323 (model claude-opus-5-5) in print mode (claude -p --permission-mode acceptEdits), on 13 November 2026 (India time), in /home/asha/projects/tuition-register (the capstone from episodes 38 and 39) and /home/asha/projects/meeras-bakes (a client brief written for this video). The agent could edit files and run only npm test, npm install, npm audit, node --test, mkdir, ls, git log and git show. Replies are quoted exactly, shortened where marked "...". Every curl request, browser test (Playwright Chromium 360 × 740), git command and calculation below was done by us.

Sources fetched on 13 November 2026: the Stack Overflow 2025 Developer Survey results post (stackoverflow.blog); Razorpay's pricing article (razorpay.com/blog/razorpay-payment-gateway-pricing-explained, updated 2 November 2026); Meta's WhatsApp Business Platform pricing page (developers.facebook.com). Prices and fees change: check the provider's own page on the day you quote.

Sunita Ma'am, Asha, Meera and the students are fictional. meera@examplebank is a made-up UPI ID.

The problem: a cook's job trial

Getting hired as a cook Getting hired to build software
Cook in front of them: a certificate proves nothing Shipped projects: used by a real person
"Why this much salt?": explain every choice The interview: explain your own code
Catering a wedding: a written quote first (menu, plates, price, date) Freelancing: scope, hours, price

The obvious move fails

"Generate 12 apps in a weekend, push them all to GitHub, write Expert in AI coding on the resume." Then the interviewer opens one file: "Explain line 47. Why this, and not that?" Silence.

Why it fails: anyone can generate code now, the interviewer too, in ten seconds. What they can't get from the AI is a person who explains it, checks it and owns it. That's what you're selling.

What employers see

From the Stack Overflow 2025 Developer Survey (49,000+ developers), as reported in its results post:

Finding Number
Use AI tools in their workflows 80%
Trust the accuracy of AI 29% (down from 40%)
Top frustration: "AI solutions that are almost right, but not quite" 45%
"Spending more time fixing 'almost-right' AI-generated code" 66%
Say vibe coding is not part of their professional work about 72%

So the skill a team pays for is catching the almost-right: exactly what this course trained in every episode.

A case study you can defend (real)

A portfolio is the set of projects you show an employer. Its best item is a case study: one page telling the story of one real project: the problem, what you decided, what went wrong and how you found it.

Prompt L:

claude -p "I'm applying for junior developer jobs. Write a README.md case study of this project for my portfolio: the problem, what I built, how I worked with an AI agent, the hardest bugs, and what I'd do next. Use git log, SPEC.md and the tests. Keep it to one page. Write it to PORTFOLIO.md."

Lines from the AI's draft:

  • "Tests are the contract ... I reviewed tests before code."
  • "Small steps ... Nothing went in with a failing test."
  • "Review before launch. A review of the finished app listed six problems ... Each problem became a test before it was fixed."
  • Hardest bug 3: "Backups that could lose data ... Restore had two subtler problems. Leftover -wal/-shm files could spoil the restored data, so restore deletes them ..."

Its reply, to be fair, warned: "'I reviewed tests before code' ... is my reading of the evidence ... Only you know if that's how it actually went", and "ask her before it goes public, or change it to 'a home maths tutor'."

Our check against git log

$ git show --stat --format="%h %s" 4175dd3
4175dd3 PIN login: 4-digit PIN from .env, sessions, 5-try pause (51 tests pass)
 ...
 src/auth.js             |  84 ++++++++++++++++++++++++++++++++
 ...
 test/auth.test.js       | 124 ++++++++++++++++++++++++++++++++++++++++++++++++
$ git show --stat --format="%h %s" de2ff1c
de2ff1c Security: HOST, Origin + Host checks, error handler, no framing (73 tests pass)
 ...
 src/security.js         |  80 +++++++++++++
 ...
 test/security.test.js   | 293 ++++++++++++++++++++++++++++++++++++++++++++++++
Claim Truth
"Each problem became a test before it was fixed" False: each fix and its tests arrived in the same commit.
"I reviewed tests before code" Never happened.
"Restore had two subtler problems" (as my hardest bug) The AI found and fixed these inside its own run. Asha never saw them, so she can't talk about them.

The rule: only claim what you can explain and prove. Every line should point to a commit, a test or a command output. Asha's rewrite (in the repo as PORTFOLIO.md):

## How I worked with an AI agent
- I wrote the spec and the acceptance criteria; the agent (Claude Code) wrote the code.
- After every step I checked it myself: a phone-size browser, sqlite3, and a real-life case.
  Then I committed (test count 29 → 41 → 51 → 73 → 88 → 101).
- I asked the agent for a security review with NO edits, then fixed each problem in its own
  prompt, with the review pasted in.
- I asked the agent to interview me on this code. It found a real gap (a wrong payment could
  not be fixed); I fixed it and checked it in the browser.

## Bugs I found myself
1. A November payment typed in on 1 December was filed as December. The AI's tests pinned the
   clock to 30 November, so they shared its assumption. Fix: a "for month" choice and AC7.
2. A guest on the Wi-Fi could lock Sunita out with 5 wrong PINs. Kept for v1, Decision 7.
3. A restart test read the laptop's real .env and hung (exit 124).
4. Rs 15000 typed instead of 1500 showed "Paid" and "Collected Rs 19000", with no way to fix it.

What goes in a portfolio

  1. 2-3 shipped projects beat 12 demos. Shipped = a real person uses it.
  2. A one-page case study for each: the problem, your decisions, your bugs.
  3. A repository (the project's folder on GitHub) with a README, small clear commits, passing tests and no secrets.
  4. Screenshots, and a demo link or a short screen recording.
  5. Say plainly that you built it with an AI agent. The point is what you decided and what you caught.

A mock interview by the AI (real)

A mock interview is a practice interview. Prompt M:

claude -p "Act as the interviewer for a junior web developer job. Read this project's code. Don't change any files. Give me 6 questions you would ask me about THIS code, from easy to hard, each pointing at a real file and line, and for each one the points a strong answer must mention. Make at least one question about a weakness in the code that I should admit."

Its six questions (shortened):

# Question Points a strong answer must mention (the AI's, shortened)
1 src/db.js:23 "If Sunita taps Present and then Absent, how many rows are there, and why?" The composite primary key makes a second row impossible; ON CONFLICT ... DO UPDATE (an "upsert": insert, or update if it exists); ? placeholders; the 303 redirect after a POST.
2 src/dates.js "Why Asia/Kolkata, why en-CA, and why add T00:00:00Z?" The server's clock may not be India time; en-CA formats YYYY-MM-DD; pinning parsing to UTC stops a date shifting by a day; for_month and paid_on are separate.
3 src/views.js "A parent's name is typed as <img src=x onerror=alert(1)>. Walk me through what happens." Escaping is the default; raw() only on purpose; quotes on attributes; escaping isn't enough inside <script> or href="javascript:...".
4 src/auth.js "Why hash the PIN before comparing it? How long would it take to guess a 4-digit PIN?" timingSafeEqual needs equal lengths; only the token's hash is stored; ~500 hours worst case; the counter resets on restart; HttpOnly, SameSite=Lax, no Secure on plain HTTP.
5 A weakness: "Sunita types Rs 15000 instead of 1500. What does she do now?" "She can't fix it. Nothing in the app can correct or remove a payment ... I'm looking for you to admit the problem plainly, not defend it." Also: a dropped connection can save a payment twice; overpaying is silent; raising a fee changes old months.
6 src/security.js "SameSite=Lax already blocks cross-site POSTs. Why do you also need the Origin and Host checks?" DNS rebinding (Host still says the attacker's name); "site" ignores the port, so a page on the same IP but another port is same-site; the Referer fallback; none of it protects plain HTTP on shared Wi-Fi.

Our check: is question 5 real?

POST /fees/4/pay  amount=15000 ...  -> 303
$ sqlite3 data/register.db "SELECT ... WHERE s.name='Om Patil'"
Om Patil|15000|2026-11-13|2026-11|cash

Fees · November 2026
Collected Rs 19000 · Still due Rs 2000     # the real total was Rs 5500
Om Patil: Paid

$ grep -nE "app\.(post|get)\(" src/app.js | grep -i pay
162:  app.get('/fees/:id/pay', ...
167:  app.post('/fees/:id/pay', ...

The payment form can only add. 88 tests passed, and not one asked "what if she types it wrong?" Our own spec never asked either.

Fees after the typo: Collected Rs 19000, Om Patil Paid

How to answer a weakness question

  1. Admit it plainly: "She can't fix a wrong payment. That's a gap."
  2. Why it happened: "Our spec never asked what happens when she types it wrong."
  3. The fix: never delete (old records stay), so cancel instead; and an "Are you sure?" check for big amounts.
  4. Best of all: "...and I fixed it. Here's the commit."

The fix (real)

Prompt O:

claude -p "A real case: Sunita typed Rs 15000 instead of 1500 for Om Patil (November). Fees now says Om is Paid and 'Collected Rs 19000', and nothing in the app can fix it. Add a 'Cancel payment' button on the Student screen that marks the payment as cancelled (never deletes it: old records stay, as the spec says), keeps it visible as cancelled, and leaves it out of every total and fee status. Also ask 'Are you sure?' before saving any payment more than 3 times the monthly fee. Add tests and run all tests. Nothing else."

Result: 100 tests pass. A new cancelled_on column; every total and fee status now has AND p.cancelled_on IS NULL. The AI's warnings: "her current register has no column for the cancel date. The app adds it the next time the server starts ... I'd run npm run backup before restarting", and "Om's record: his Rs 15000 is not fixed yet ... press Cancel payment on the Rs 15000, then record the correct Rs 1500."

The key change in src/db.js (the AI's code):

// Registers made before "Cancel payment" have no cancelled_on column: add it (old payments stay counted).
const cols = db.prepare('PRAGMA table_info(payments)').all().map((c) => c.name);
if (!cols.includes('cancelled_on')) db.exec('ALTER TABLE payments ADD COLUMN cancelled_on TEXT');
...
LEFT JOIN payments p ON p.student_id = s.id AND p.for_month = @month AND p.cancelled_on IS NULL
...
cancelPayment: db.prepare('UPDATE payments SET cancelled_on = ? WHERE id = ? AND cancelled_on IS NULL'),

Changing the shape of a database that already holds data is called a migration: here, adding one column that old rows leave empty.

Our check (backup first, then the browser):

$ npm run backup
Saved backup: backups/register-2026-11-13_09-53-11.db
# Chromium, 360 x 740: Om -> Cancel payment, then record Rs 1500
1 payment · Rs 1500 in all · 1 cancelled (not counted)
Cancelled on Fri, 13 Nov, 2026 · not counted
Fees: Collected Rs 5500 · Still due Rs 2000
Om Patil|15000|2026-11-13|2026-11|cash|2026-11-13
Om Patil|1500|2026-11-13|2026-11|cash|
# Aditya: type 5000 -> "Are you sure? ..." -> she corrects it to 500
buttons: [ 'Yes, save Rs 5000' ]      # the only button; it saves 500

A UX bug (user experience: how it feels to use) that no test caught: the only button still said "Yes, save Rs 5000" after she corrected the amount.

Before: Rs 500 typed, the button says Yes, save Rs 5000

Prompt P:

claude -p "Checked in Chromium: after the 'Are you sure?' warning for Rs 5000, Sunita corrects the amount to 500, but the only button still says 'Yes, save Rs 5000' (it saves 500). Fix only that: after the warning, show two buttons: 'Save' (normal check again) and 'Yes, save Rs 5000' that only saves if the amount is still the one she was warned about. Add a test, run all tests. Nothing else."

101 tests pass. It also found a trap: "In browsers, a disabled button's value isn't sent, so that would have dropped the Yes button's amount ... This is the part only a real browser can confirm." Our Chromium check:

after warning, buttons: [ 'Save', 'Yes, save Rs 5000' ]
500 + Yes  -> "The amount is now Rs 500, not Rs 5000. Press Save payment to save Rs 500."
5000 + Save -> "Are you sure? Rs 5000 is more than 3 times the monthly fee (Rs 1500). ..."
5000 + Yes -> saved: Meher Gill|5000|upi

Now question 5 has the best possible answer: found in a mock interview, admitted, fixed with tests, checked by hand.

After: two buttons

Om's cancelled payment, kept and not counted

Your first freelance quote (real)

  • Freelancer: paid per project, by many clients, instead of a salary from one employer.
  • Brief: what the client asks for, in their own words.
  • Scope: exactly what's included, and what's not.
  • Quote: scope + hours + price + timeline + payment terms. Write it before any code.

The brief (BRIEF.md, written for this video):

Hi! I run Meera's Bakes from home in Pune. I want a website where
customers can see my cakes, choose a date and order, and pay online
by UPI. I should get a WhatsApp message for every order. Can you make
it in 2 weeks? My budget is Rs 15,000.

Missing from it: how many cakes, delivery or pickup, what happens when a date is full. A brief is a draft of the scope, never the scope.

Prompt N:

claude -p "A client sent me the message in BRIEF.md. I'm a beginner freelancer who builds with AI agents. Write me a fixed-price quote I can send her, in rupees: the tasks with hours for each, the total, what's included and the timeline. Write it to QUOTE.md."

The AI's quote (shortened):

Payment Hours Price
Option A (recommended) UPI through Razorpay (a payment gateway: a service that takes online payments for you); orders confirm automatically 40 Rs 18,000
Option B A UPI pay link; Meera confirms each payment herself 34 Rs 15,000

Tasks (Option A): kick-off call 2 h · design and pages 6 · menu 2 · order form and date rules 5 · Razorpay UPI payment 6 · WhatsApp alert + email backup 5 · Google Sheet of orders 3 · policy pages 2 · domain and hosting 2 · phone testing 3 · two change rounds 2 · handover video 2. Payment terms: 50% advance, 50% on launch. Not included: running costs ("Razorpay fee ... about 2% per online payment", "WhatsApp messages ... usually under Rs 1 per order"), photography, logins, more than 20 cakes at launch; later changes at Rs 500/hour. It said: "I priced it at about Rs 450 an hour ... I didn't just shrink everything to fit her Rs 15,000", and "The running costs are estimates ... Check current prices quickly before you send."

Our three checks

# 1. the maths
40 h x Rs 450 = Rs 18,000     (Option A: matches)
34 h x Rs 450 = Rs 15,300     (Option B: priced Rs 15,000 = her budget)

# 2. the gateway fee: Razorpay's own pricing article (fetched 13 Nov 2026)
"UPI: 2%" platform fee, and "The platform fee attracts 18% GST"
2% + 18% of 2% = 2.36%   ->  on a Rs 1,000 cake: Rs 23.60

# 3. Option B's UPI link, opened on a LAPTOP (Chromium)
<a href="upi://pay?pa=meera@examplebank&am=1000&cu=INR">Pay</a>
click  -> nothing happens (page stays the same)
goto   -> net::ERR_ABORTED at upi://pay?pa=meera@examplebank&am=1000&cu=INR
  1. Option B was quietly fitted to the budget (Rs 300 under 34 × 450). That's fine, but know it's a discount.
  2. "About 2%" left out tax. GST is India's goods and services tax; it applies to Razorpay's fee, so the real cost is 2.36%. (The same article describes a 0% platform-fee offer for eligible new merchants for up to 90 days; terms apply, so check the current page.)
  3. A upi:// link opens a UPI app on a phone. A laptop has no UPI app, so nothing happens: Option B also needs a QR code (a square barcode a phone camera can scan to pay).

The WhatsApp cost: Meta's pricing page says that since 1 July 2025 it charges per message, by message category (marketing, utility, authentication) and country, and that "utility templates delivered within an open customer service window are free". The rates are in Meta's rate cards, so look up the current India utility rate before you promise a number.

Pricing in plain words

  • Hourly: paid for your time. Fixed price: paid for the result, however long it takes you.
  • Fixed price = your hours estimate × your rate + a buffer. A buffer is extra time: beginners always guess low, and testing and changes take longer than coding.
  • Scope creep: "just one more small thing", twenty times. Defences: a fixed number of change rounds, an advance payment, and a "not included" list.
  • Faster isn't cheaper. With an AI agent you build faster; the client still pays for a working shop that takes orders, not for your typing.

Where the first paid work comes from

  1. People you know: a tutor, a shop, a clinic, a housing society office.
  2. Freelance platforms (websites that connect clients and freelancers): their fees change, so read the platform's current fee page before you set a price.
  3. Open source (projects whose code is public): fixing small issues gives you real code reviews to show.
  4. Your case studies, shared where your future clients are.

Two nevers: never paste a client's passwords or API keys into a prompt, and never take work you can't support after launch.

What to learn next

  • One language, deeply: JavaScript or Python.
  • DSA (data structures and algorithms): how to store data and solve problems efficiently; most interviews test it.
  • System design: how big apps fit together.
  • SQL and databases, Git, security basics, testing.

All of these are free full courses on SyntaxPilot. The better you know code, the more the AI can do for you, because you can check it.

Checklist

  • [ ] 2-3 shipped projects, each with a real user
  • [ ] A case study per project: every claim has proof
  • [ ] Ask the AI to interview you on your own code
  • [ ] Find a weakness, admit it, fix it, commit it
  • [ ] Quotes: scope, not-included list, hours × rate + buffer
  • [ ] Every price and fee from the provider's own page
  • [ ] Keep learning: a language, DSA, system design

Common mistakes (real)

Mistake What we saw Why Fix
12 AI repos, no story "Explain line 47." ...silence Generated, not owned 2-3 shipped projects + case studies
An AI-written case study sent as is "Each problem became a test before it was fixed" False claims Check every line against git log
Defending a weakness "Collected Rs 19000" (really Rs 5500) Hides the gap Admit it, fix it, commit it
Fees in a quote taken from the AI "about 2%" (really 2.36% with GST) Prices change; tax left out The provider's own page
A UPI link only net::ERR_ABORTED on a laptop No UPI app on a laptop Add a QR code

Mini project

  1. Ask your agent for a portfolio case study of your capstone. Mark every sentence: proved (commit/test/output), not proved, or false. Rewrite.
  2. Ask it to interview you on your code (6 questions, real files and lines, one weakness). Answer out loud, then fix the weakness with one prompt and check it in a browser.
  3. Write a quote for a real business near you: scope, not-included, hours per task, rate, buffer, payment terms. Check every fee on the provider's page and write the date you checked.

Practice questions

  1. Why is "12 AI-generated repos" a weak portfolio?
Answer

Anyone can generate code. Without a real user, a case study and the ability to explain the code, the repos show nothing the interviewer couldn't produce in seconds.

  1. Your AI-written case study says "I wrote tests before every fix." How do you check it?
Answer

git show --stat <commit> for each fix: if the test file and the code arrive in the same commit (or the test comes later), the claim is false. Rewrite it to what the history proves.

  1. An interviewer finds a real bug in your project. Give the four-part answer.
Answer

Admit it plainly; say why it happened; describe the fix; best of all, show the commit that fixed it.

  1. 34 hours at Rs 450/hour with a 20% buffer: what fixed price?
Answer

34 × 1.2 = 40.8 hours ≈ 41 hours × Rs 450 = Rs 18,450 (or 34 × 450 = 15,300 × 1.2 = Rs 18,360). Either way, about Rs 18,400.

  1. A gateway charges a 2% fee plus 18% GST on that fee. What does a Rs 2,500 order cost in fees?
Answer

2% of 2,500 = Rs 50; GST 18% of 50 = Rs 9; total Rs 59 (2.36%).

Quiz

The interviewer points at a bug in your project you haven't fixed. Best answer?

  • A) "That's not a bug, it's how it's designed."
  • B) "Yes, that's a gap. It happened because... I'd fix it by... (or: here's the commit)."
  • C) "The AI wrote that part."
  • D) Change the subject to a feature that works
Answer

B. Admit, explain, fix. A defends it, C blames the tool you chose to use, D avoids it.

Interview questions

  1. "Did you use AI to build this?" Yes: Claude Code wrote most of the code. I wrote the spec and acceptance criteria, ran every step's checks myself (browser, sqlite3, real-life cases) and committed after each one. Here are three bugs I caught that its tests missed.
  2. "Walk me through how you'd make a payment uncorrectable mistake fixable." Never delete financial records; add a cancelled_on column (a migration), exclude cancelled rows from every total, keep them visible, and add a confirmation for unusual amounts.
  3. "How do you price a fixed-price project?" Break it into tasks with hours, add a buffer, multiply by my rate, list what's not included, limit change rounds, and take an advance.

Glossary

Term Meaning (with an example)
Portfolio The set of projects you show an employer or client.
Case study A one-page story of one project: problem, decisions, bugs, results.
Shipped In real use by a real person (Sunita's register).
Repository (repo) A project's folder with its full Git history, e.g. on GitHub.
README The first file people read in a repository: what it is and how to run it.
git show --stat Shows which files a commit changed and by how much.
Mock interview A practice interview; here the AI plays the interviewer.
Upsert Insert a row, or update it if it already exists (ON CONFLICT ... DO UPDATE).
Migration Changing the shape of a database that already has data (adding cancelled_on).
UX (user experience) How it feels to use something; a button label that lies is a UX bug.
Freelancer Paid per project by many clients, not a salary from one employer.
Brief What the client asks for, in their own words.
Scope Exactly what's included in the work, and what's not.
Quote Scope + hours + price + timeline + payment terms, agreed before work starts.
Fixed price / hourly Paid for the result / paid for time.
Buffer Extra time added to an estimate because estimates run low.
Scope creep Small extra requests that add up to a much bigger job.
Advance payment Part of the price paid before work starts (e.g. 50%).
Payment gateway A service that takes online payments for a website (Razorpay).
UPI India's Unified Payments Interface: instant bank-to-bank payments from a phone app.
upi:// link A link that opens a UPI app with the payee and amount filled in; needs a UPI app.
QR code A square barcode a phone camera scans, e.g. to pay by UPI from a laptop screen.
GST India's goods and services tax; 18% on the gateway's fee.
Platform fee What the gateway charges per payment (2% on UPI in Razorpay's article).
Freelance platform A website that connects clients and freelancers, and usually takes a fee.
Open source Software whose code is public, so anyone can read it and suggest fixes.
DSA Data structures and algorithms: storing data and solving problems efficiently.
System design Planning how the parts of a large app fit together.

Summary

  • Generated code is cheap; explaining, checking and owning it is the skill employers pay for (80% use AI, 29% trust its accuracy).
  • The AI's case study read well but had claims Asha couldn't defend. git show --stat proved one false. Only claim what you can explain and prove.
  • The AI's mock interview found a real weakness: a typo of Rs 15000 showed "Paid" and "Collected Rs 19000", with no way to fix it, and 88 tests green. We admitted it, fixed it (cancel, never delete; "Are you sure?"), and caught a lying button label in the browser.
  • The AI's freelance quote was thoughtful, but: Option B was quietly fitted to the budget, the gateway fee left out GST (2.36%, not 2%), and a UPI link does nothing on a laptop.
  • Price fixed jobs as hours × rate + buffer, with scope, a not-included list, change rounds and an advance.
  • Next: one language deeply, DSA, system design. That's the end of the Vibe Coding course: keep building, and keep checking.