Engineeringhot

Vibe Coder

Also hired as: AI-Native Developer · Product Engineer (AI-assisted) · Solo Builder

Direct AI agents that write the code. You own the spec, the taste, and the ship button.

US salary · 2026
$120k – $260k+ total comp as an employee; uncapped (and unfloored) solo
You'll rarely see 'vibe coder' on an offer letter — comp follows product-engineer bands at whatever company hires you, equity-heavy at startups. The solo path trades salary for ownership and is a real path now, not a meme.
Typical background
Self-taught builders, product-minded engineers, and designers or PMs who crossed over once typing stopped being the bottleneck.
This roadmap
5 stages · 25 nodes
5 proof-of-work checkpoints
What is a Vibe Coder?

A vibe coder builds and ships real software by directing AI coding agents — Claude Code, Codex, Cursor, Antigravity — instead of typing most of the code by hand. Andrej Karpathy coined the term in early 2025, half-jokingly, for coding without reading the diffs; by 2026 it names a serious discipline with the joke removed. The craft moved up a level: you write specifications, hold architectural taste, verify relentlessly, and orchestrate several agents in parallel. The agent owns the keystrokes. You own everything that matters — what gets built, whether it's right, and whether it ships.

The role exists because agentic coding crossed a threshold: from autocomplete to delegation. Teams discovered a brutal spread between people who use these tools casually and people who run them well — same tools, wildly different output. So companies began hiring for demonstrated agent leverage directly: 'AI-native developer' and product-engineer listings now ask how you work with agents, not whether. Meanwhile the solo path became real. One person with good taste and a fleet of agents can ship what used to take a team — and either sell it as a product or sell the capability as a service.

It is not prompt engineering — you ship products, not prompts. And it is not just 'software engineer, but faster.' The unit of work changes from the function you write to the task you delegate, so the scarce skills change too: writing specs an agent can't misread, reviewing diffs fast without rubber-stamping, building verification harnesses so agents check themselves, and knowing precisely when to grab the keyboard. People with all four skills are still rare, which is the whole opportunity. This roadmap builds them in order.

What you'll actually do
  • Turn fuzzy product ideas into written specs an agent can execute: acceptance criteria, edge cases, and explicit non-goals.
  • Run two to five agent sessions in parallel across git worktrees, slicing features into agent-sized tasks that don't collide.
  • Review AI-written diffs fast and honestly — read every line that touches auth, money, or data deletion; skim the rest with tests as backstop.
  • Maintain an agent-legible codebase: CLAUDE.md conventions, memory files, and docs that make every next session smarter.
  • Build the verification harness — types, tests, lint, CI — and wire agents to run it themselves before showing you anything.
  • Ship to production and stay there: monitoring, error triage, and fixes with the agent in the loop.
  • Re-evaluate the tool stack monthly; models and harnesses move fast enough that last quarter's workflow is stale.
  • Drop to hand-written code when the agent is flailing — and recognize that moment early instead of burning three more sessions.
This role fits you if
  • You think in outcomes and constraints, not syntax; writing a crisp spec feels like the real work.
  • You'd rather review three diffs than hand-write one function.
  • You can smell plausible-looking slop — code that compiles, skims well, and is still wrong.
  • You ship side projects habitually, and you finish them.
  • You have the patience to verify: nothing merges because 'the agent seemed confident.'
The toolbox
Claude CodeCodexCursorAntigravityGit worktreesCLAUDE.md / agent memoryTypeScriptNext.jsSupabasePlaywrightGitHub ActionsMCP serversVercelCode review
The roadmap — 5 stages, 25 nodes
Stage 1

Foundations: code + agents

Weeks 1–4

You can't direct what you can't judge. Learn enough code to read it critically, then get your first agent shipping under your direction.

Learn to read code, even if you rarely write itCore
JavaScript/TypeScript basics, git, what a PR is, how a web app is shaped. The bar isn't writing from scratch — it's looking at a diff and knowing what it does and what it risks.
First agent sessionsCore
Install Claude Code and learn the loop: describe, watch, correct, repeat. Get a feel for what one session can and can't hold.
How LLMs work — enough to debug themCore
Context windows explain most agent failures: the model didn't 'forget', the context overflowed. This mental model turns mysterious behavior into fixable behavior.
Try a second tool earlyElective
Spend a weekend with Codex or Antigravity to see which habits transfer and which are tool quirks. You're learning the category, not memorizing one product.
Checkpoint: ship an app entirely through an agent✓ Checkpoint
A real web app — auth, database, deploy — live on a public URL, built end to end by directing an agent. Then fix one production bug after launch. Shipping plus one repair is the full loop.
Stage 2

Specification and memory

Weeks 5–8

The core law of the craft: bad agent output usually means your spec was ambiguous. Learn to write instructions that survive contact.

Specs an agent can't misreadCore
Acceptance criteria, edge cases, explicit non-goals, and examples for the hard cases. When output disappoints, fix the spec before blaming the model — that reflex is the skill.
CLAUDE.md and project memoryCore
Encode conventions once — stack choices, naming, forbidden patterns — so every session starts smart. Your memory file is compound interest on every future task.
Context engineeringCore
Feed the right files, keep sessions scoped to one concern, and restart when context rots. Knowing when a session is poisoned saves hours of correction loops.
Prompting patterns for codeCore
Plan-then-execute, test-first prompts, and asking the agent to state assumptions before building. Structured prompting is the difference between a junior and a senior agent.
Checkpoint: the cold-handoff spec✓ Checkpoint
Write a one-page spec for a nontrivial feature. Hand it to a fresh agent session with zero back-and-forth beyond the document. The diff should land with at most two correction rounds — if it takes more, the spec failed, not the agent.
Stage 3

Verification and taste

Weeks 9–14

Speed without verification is how vibe-coded products die in public. Build the review skill and the harness that makes speed safe.

Review AI diffs fastCore
Always read: auth, payments, data deletion, migrations, and anything touching secrets. Learn to skim the rest against tests. Rubber-stamping is the failure mode that ends careers.
Build the verification harnessCore
Types, tests, lint, and CI wired so the agent runs them itself and iterates until green. The harness is what lets you safely not read every line.
Develop architectural tasteCore
Read excellent open-source codebases and learn the patterns by name. Taste is what keeps a codebase coherent across fifty agent PRs — without it, entropy wins by PR twenty.
Security review habitElective
Agents introduce injection risks, leaked secrets, and over-permissive endpoints at machine speed. Build a pre-ship checklist and actually run it.
Checkpoint: the defect hunt✓ Checkpoint
Take an agent-built codebase (your Stage 1 app works) and find five real defects the agent introduced — logic, security, or data bugs, not style nits. Document each with root cause and the spec or harness change that would have prevented it.
Stage 4

Orchestration and scale

Weeks 15–20

One session is a skill; a fleet is a system. Parallelize safely, delegate to subagents, and keep cost per feature visible.

Parallel sessions with worktreesCore
Run independent tasks in separate git worktrees so agents can't stomp each other. Learn which work parallelizes cleanly and which serializes on your review bandwidth.
Subagents and pipelinesCore
Fan work out to specialized subagents, fan results back in, and automate the repeatable paths into PR pipelines.
MCP: give agents your toolsCore
Wire agents to your database, browser, and deploy pipeline through MCP servers so they can verify their own work end to end instead of guessing.
Cost and model disciplineCore
Track spend per task, route cheap models to cheap work, and know your cost per shipped feature. Orchestration without budgets is how a side project develops a cloud bill.
Checkpoint: the one-week parallel release✓ Checkpoint
Ship a multi-feature release in one week using three or more parallel agent streams. Keep a log: prompts, correction rounds, and cost per feature. The log is the proof — and your portfolio piece.
Stage 5

Ship, prove, get paid

Weeks 21–28

The market pays for shipped products and legible process. Get real users, publish how you work, and pick your monetization path.

Production operationsCore
Monitoring, error tracking, and incident fixes with the agent in the loop. 'I operate what I ship' is what separates you from the demo crowd.
Public proof-of-workCore
Publish build logs: what you asked, what the agent did, what you corrected, what it cost. In 2026 hiring, a legible process is worth more than a résumé line.
Pick your laneCore
Employee (product-engineer roles that screen for agent leverage), agency/freelance (sell speed), or solo products (sell outcomes). Each wants different proof — target one deliberately.
Speed repsElective
Timed rebuilds of small apps keep your delegation sharp the way typing tests never did. Treat it as training, not performance.
Checkpoint: real users or real revenue✓ Checkpoint
Ship a product to ten external users or the first dollar of revenue, and publish the build log showing agents did the majority of the code. That artifact — shipped product plus legible process — is the whole job application.
Build your portfolio

Nobody hires a Vibe Coder off a certificate. They hire off proof. Ship these and put them where people can click them:

01
The agent-built app — and its first repair
A real web app with auth, a database, and a deploy, built end to end by directing an agent — plus one production bug found and fixed after launch. Live URL, public repo.
Proves: You can run the full loop — ship, break, repair — not just demo.
02
The cold-handoff spec
A one-page spec for a nontrivial feature, published side by side with the diff a fresh agent session produced from it in at most two correction rounds.
Proves: Your specs are precise enough to direct a workforce that can't read your mind.
03
The defect hunt
Five real defects — logic, security, or data — found in an agent-built codebase, each documented with root cause and the spec or harness change that would have prevented it.
Proves: You verify instead of rubber-stamping, which is the failure mode that ends vibe-coding careers.
04
The one-week parallel release
A multi-feature release shipped in one week using three or more parallel agent streams in git worktrees, with a log of prompts, correction rounds, and cost per feature.
Proves: You can orchestrate a fleet, not just drive one session.
05
A product with users and a public build log
Something shipped to ten external users or the first dollar of revenue, with a published build log showing agents wrote the majority of the code — what you asked, what you corrected, what it cost.
Proves: Shipped product plus legible process — the whole job application in one artifact.
Position your profile
Headline formula

I ship production software by directing AI agents — [N] products live, [X]% agent-written, verified by a harness I built.

Resume bullets to earn
  • Shipped [product] to [N] users solo in [X] weeks by running three parallel agent streams; [Y]% of merged code was agent-written.
  • Cut feature cycle time from days to hours with cold-handoff specs that land in at most two correction rounds.
  • Built a verification harness — types, tests, lint, CI — that agents run themselves; [N] releases with zero rollbacks.
  • Found and documented [N] agent-introduced defects pre-ship, each traced to a spec or harness fix.
  • Held cost per shipped feature to $[X] across [N] features by routing models and scoping sessions deliberately.
Where to be visible
  • Build in public with real logs — prompts, corrections, costs — not hype threads. Legible process is the differentiator; everyone else posts screenshots.
  • Pin the product repo, its spec, and its CLAUDE.md together on GitHub; the README leads with what it does and who uses it.
  • Publish one honest failure teardown for every win — the defect-hunt post outperforms the launch post with hiring managers.
  • Ship where builders watch: demo days, launch platforms, and the communities around your agent stack.
  • Contribute agent-workflow writeups or small PRs to the tooling you use — the ecosystems are young enough that one good contribution gets noticed.
What interviews actually test
  • A live agent-driven build while they watch: spec it, delegate it, review the diff out loud. They're grading your corrections, not the agent's code.
  • A diff-review round with planted bugs. Read auth, money, and deletion paths line by line — that habit is the pass signal.
  • A process deep-dive on your build log: what did you decide versus the agent? Own the architecture calls explicitly.
  • System design framed as delegation: what you hand the agent, what you write yourself, and why.
  • On the solo path there is no interview — the build log and the revenue are the pitch. Package them like it.
Who's hiring
Seed-to-Series-B startups (every YC batch since 2024 assumes it)ShopifyVercelReplitAnthropicOpenAIProduct agencies shipping client work with three-person teamsYour own company — the solo path is real
Free courses that map to this role
Related roles
FAQ

Is vibe coding a real job or just a meme?

Both, which confuses people. The meme is coding without reading the output. The job — posted as AI-native developer or product engineer — is directing agents with rigorous specs and verification, and companies now screen for it explicitly. There's also a real solo path: one person shipping revenue-generating products with agent leverage.

Do I still need to learn to code?

You need to read code and judge it, which is a smaller but non-optional subset of learning to code. Skipping it caps you at demos: the first production bug, security review, or gnarly migration ends the run. The good news is you can learn to read critically much faster than you can learn to write fluently.

Won't this job be automated away too?

The typing already was — that's the premise. What remains is deciding what to build, specifying it, and verifying it, and those move up in value as execution gets cheap. If agents improve tenfold, the person who directs them well ships tenfold more. The role transforms; it doesn't vanish.

How is a vibe coder different from a prompt engineer?

A prompt engineer optimizes model interactions that live inside a product. A vibe coder ships whole products, using agents as the workforce. The overlap is prompting skill; the difference is that your deliverable is working software in production, with everything that entails — architecture, review, ops.

Which tool should I learn first?

Pick one deep rather than three shallow — Claude Code is the most common choice for serious work and this site's core track. The concepts (specs, memory, verification, orchestration) transfer to Codex, Cursor, and Antigravity almost 1:1. Add a second tool in month two to prove that to yourself.

Start building, today

Every stage above maps to free lessons on this site. No signup, no paywall — open the first course and ship your first checkpoint this week.

Browse the courses →