Skip to content

Filling Out an Amazon Job Application — Plain-Language Notes

WorkWingman applies to jobs for you. Amazon runs its own application system at amazon.jobs — they built it themselves rather than buying a system from a company like Workday or Lever. To learn how to fill it out reliably — without ever touching the real site — we built a practice range, the same way we did for Workday and Lever.

The practice range (and why it's safe)

We built a fake Amazon application form that only lives on this computer. It runs on 127.0.0.1, the address a computer uses to talk to itself. Nothing goes out to the internet. There is:

  • no real amazon.jobs website involved,
  • no real Amazon account — the "sign in" step you see is completely made up locally; it never checks a password against anything real, and no account is ever created anywhere,
  • no Submit that does anything — the Submit button is a decoy that just writes a note in the log ("would submit — suppressed") and starts over.

Then we let the robot practice on this fake form hundreds of times, each time with the boxes shuffled and renamed a little, and we wrote down what worked and what didn't.

What an Amazon application is like

Think of an Amazon application as a series of screens you click "Continue" through:

  1. Sign in or create an account (sometimes skipped, if you're already "logged in").
  2. Your name, email, phone, and address.
  3. A place to attach your resume (in the practice range we never actually attach a file).
  4. A handful of job-specific questions — "Are you allowed to work here?", "Do you need a visa?", "When can you start?", and sometimes softer ones like "Why do you want to work at Amazon?" or "Which shift do you want?"
  5. Some optional demographic questions (gender, ethnicity, veteran status, disability) — entirely voluntary, never required.
  6. A Review and Submit screen at the end.

Why Amazon is the trickiest practice range so far

Workday and Lever are each built by one company and sold to thousands of employers, so every Workday form and every Lever form is labeled the same underlying way, even if it looks different on the surface. Amazon built its own system in-house, with different internal teams working on different parts of it over the years. That means the labels under the hood are much less consistent — sometimes a box has a clean, obvious label; sometimes it's been renamed by whichever team touched it last; sometimes there's no computer-readable label at all and the only way to find the box is to read the human-readable text next to it and figure out which box that text belongs to.

That inconsistency is exactly why this is a good practice range: it's the closest thing we have to "a company that didn't buy someone else's form-builder," and it's a preview of the kind of messiness the robot will run into on lots of smaller companies' custom career sites too.

How the robot copes with backup ways to find each box

For every box, the robot has a ranked list of ways to find it — most reliable first, then backups:

  1. Guess the obvious label first (a clean, sensible id like "first name").
  2. If that's not there, look for anything with a similar-sounding label (maybe it got renamed to something like "candidate-first-name-field").
  3. If that's not there either, check for an accessibility tag (screen-reader labels are often still present even when the technical label isn't).
  4. If NONE of that works, read the actual words printed next to the box (the human-readable label) and find the box sitting right next to those words.
  5. Last resort: guess from the greyed-out hint text inside the empty box.

Across a full practice run — 20 applications, each with different boxes renamed or hidden — the robot completed every single one with zero crashes, using its backup methods whenever the first guess didn't pan out.

When the robot stops and asks

Some questions can't be answered safely from your saved profile — like "Why do you want to work at Amazon?" (that's genuinely personal — the robot won't make up an answer) or "Which shift do you want?" (a soft preference, not a fact). For those, the robot pauses and asks a human rather than guessing. In the practice run, it paused for 15 out of 20 applications. The questions it can answer confidently — "Are you authorized to work here?" and "Do you need visa sponsorship?" — it fills in directly from your profile, because those have clear factual answers.

Two real bugs the practice range caught

  1. Clicking the wrong "Continue" button. The first time we ran the practice range, every single application failed. It turned out the multi-step form keeps every screen's buttons in the page at once — it just hides the ones you're not currently looking at. The robot was accidentally trying to click a "Continue" button on a hidden screen instead of the one actually showing. We fixed it so the robot only ever clicks buttons on the screen currently in front of it.
  2. Getting stuck on optional questions. The voluntary demographic questions (gender, ethnicity, etc.) sometimes offer different wording than what's saved in your profile — that's normal and fine, since they're optional. But the robot was stubbornly retrying for several seconds trying to force a match before giving up, instead of just checking first and moving on. We fixed it to check the available choices before attempting to pick one.

Both were caught specifically because we watched a full, real practice run instead of assuming the robot's logic would work first try — exactly why this practice range exists.

The bottom line

  • The robot practices on a fake, local, offline Amazon-style application only.
  • It never creates a real Amazon account, never submits a real application, and never talks to the internet.
  • It learned how Amazon's homegrown form differs from the "bought-off-the-shelf" forms (Workday, Lever), found and fixed two real bugs by watching a full practice run, and confirmed its backup-clue system works even on a form this inconsistently labeled.

Why this matters for teaching an AI to "just look at the page"

Right now the robot relies on a list of hardcoded clues we wrote ahead of time for each box. That works well here because we could study the real amazon.jobs form and write good guesses. But Amazon's form is a preview of a much bigger problem: thousands of smaller companies run their own homegrown career sites, each labeled differently, and nobody can write a hardcoded clue list for all of them in advance. That's the case for eventually letting an AI look at the rendered page itself — read the text, the layout, the labels — and figure out which box is which on the fly, the same way a person would, instead of relying on a pre-written list of guesses. This practice range's "pause and ask a human" moment is designed so that kind of smarter, look-and-figure-it-out approach could slot in later without changing how the rest of the system works.