An AI job-application engine that won't make things up
I conceived, specified and directed GetMeAJob, an AI application that helps a job seeker produce strong, tailored applications without inventing anything. Claude designed the architecture and wrote the code.
This page covers the problem, the product decisions I made, what happened when I used it on my own job search, and where it stands today.
The problem
A serious job application takes real work: analyzing the role, researching the company, tailoring a resume, writing a cover letter and answering screening questions. AI can make that faster, but generic tools tend to produce generic material, and sometimes they invent experience the applicant doesn't have.
When I wrote the requirements, I set out the problems the product had to solve:
- Fabrication. The tool must never invent experience, skills, metrics, titles or dates.
- Scattered information. Several resumes can describe the same person differently, and the tool has to reconcile them.
- Optimizing for software instead of people. Tuning a resume for applicant tracking systems (ATS) shouldn't make it worse for the recruiter reading it.
- Wasted effort. Sometimes the honest answer is "don't apply," and the tool should be willing to say so.
Why I started it
I started GetMeAJob while navigating my own job search after moving to Chicago. I realized that creating a strong application involved much more than rewriting a resume. It meant understanding the company, assessing whether a role genuinely suited my experience, and presenting that experience accurately.
AI could help with that process, but I didn't want it inventing qualifications or making decisions on my behalf. So I set out to design a tool that prioritizes evidence, transparency and human judgment.
My role and Claude's
I defined what the product is and the rules it must follow. Claude decided how to build it and built it.
| Area | Me | Claude |
|---|---|---|
| Problem and vision | Identified the problem and defined the product as an AI career agent, not a resume builder | None |
| Requirements | Wrote the requirements document: 34 requirement areas, the minimum product and a five-phase plan | None |
| Product rules | The evidence-integrity rule, four evidence categories, human approval of everything, never submitting on the applicant's behalf, respecting job sites' terms, and permission to say "don't apply" | Turned these rules into enforceable mechanisms: evidence IDs, a validator, an approval gate and guardrails |
| Workflow | Defined the core journey, from uploading resumes to a tracked, approved application | Designed the data model and the processing pipeline behind it |
| Architecture and code | Set constraints: separate data for each user, no dependence on a single AI provider, security priorities | Chose the technology stack, designed the architecture and wrote all of the source code and automated tests |
| Testing in practice | Ran it on my own job search, which surfaced a real limit in the AI provider's API | Fixed the resulting test-compatibility problem and added a regression test |
The application is written in TypeScript with React, Node.js and SQLite. That code is Claude's work, so I don't present it as evidence of my own programming skills.
How it works
The applicant uploads their resumes once. GetMeAJob merges them into a single profile, called the Career Brain, in which every fact keeps a link to the document or answer it came from. Each new job then runs through the same pipeline:
- Job intelligencePaste a job description, or import it from a supported job board link. The role is broken into numbered requirements.
- Company researchWeb research with citations, sorted into facts, interpretations and hypotheses.
- Fit analysisEach requirement is matched to evidence. The AI judges each match; code calculates the scores.Can stop here with "do not apply"
- StrategyHow to position the applicant, what to lead with and what to play down, before any document is written.
- DocumentsResume, cover letter and answers to the application questions, all generated from the strategy.
- Integrity checkEvery number and tool is checked against cited evidence. A failing line gets one repair attempt, then it's removed and reported.
- Review and approvalThe applicant reviews, edits and approves everything. Nothing is ever submitted automatically.
Drawn from the project's own architecture documentation.
Four principles behind the design
Strategy before documents
Documents follow a stored strategy, not the raw job description.
Evidence over prose
Every claim cites an evidence item, and code checks it against the source.
Code where it can be certain
Names, titles, employers, dates and scores come from records and calculations. The AI chooses and phrases.
The applicant approves
Everything is a draft until a person approves it, and the app never submits anything.
Designing for trust
The hardest product question wasn't what the AI should write. It was how to stop it writing things that aren't true, without making the tool useless. The answer was to treat evidence as a first-class object:
- Each fact is labeled verified, provided by you, inferred or unsupported. Only the first two can appear in an application.
- When two resumes disagree, the app asks the applicant instead of guessing.
- Generated and edited text is checked for numbers and known tools that the evidence doesn't support. Approval is blocked while problems remain.
- Answers about work authorization, salary or legal matters need an explicit "I checked this" confirmation.
Screenshots show a fictional applicant and simulated AI output. Any scores shown come from test data, not real-world results.
Judgment before documents
Most AI writing tools start drafting immediately. GetMeAJob does the thinking first, the way a good recruiter or career coach would:
- An honest fit analysis. Requirement-by-requirement matching with cited evidence, plus separate recruiter, hiring-manager and ATS views. Guardrails, such as a role that can't offer sponsorship the applicant needs, stop the process with "do not apply."
- Research with sources. Each finding about the employer is labeled as fact or interpretation and linked to where it came from.
- A strategy. The positioning, what to emphasize, likely objections, and which keywords can't honestly be claimed.
Screenshots show a fictional applicant and simulated AI output. Any scores shown come from test data, not real-world results.
What happened when I used it for real
On October 7 and 8, 2026, I ran GetMeAJob on my own job search, using the Anthropic API and my real resumes. It merged 8 resumes into one profile, I answered 40 follow-up questions, and two real applications went all the way from research to approved, exported documents. The app's own log records 70 successful AI calls, which it estimated cost about $5.
It also hit a genuine limit. The API rejected the largest output formats (for resume extraction, job analysis and strategy) as too complex. Five steps were switched to describing the required format in the prompt instead, while still validating every response, and every later call for those steps succeeded. That change briefly broke the automated tests; it was fixed on October 26, along with a new test that fails if the workaround is ever removed.
What this doesn't show. I haven't yet evaluated the quality of the documents it produced, and nobody else has used it. The screenshots on this page come from a separate test run with a fictional applicant and simulated AI responses, not from my live applications.
Status and limitations
Phase 1, the core application engine, is built and runs on my computer. As of October 26, 2026, all 23 of its automated tests pass. Four more phases are planned but not built: job discovery, interview preparation, assisted form-filling with approval, and learning from outcomes.
Limitations, stated plainly
- The quality of its live output hasn't been evaluated.
- It's a single-user prototype that runs locally. It isn't deployed or hosted, and it isn't ready to serve other users.
- The automatic tool check uses a fixed list of about 60 common tools, so an unlisted tool could slip through.
- No live AI test has been run since a library upgrade on October 26. The automated tests, which use simulated AI responses, pass.