Career That Lasts: shipping a live AI career app with Claude Code

## the problem

## what I built

## how Claude was actually used

1. 01

   Planning before code

   Before the first line of app code, I had Claude Code write the plan: a product brief, sitemap, database schema, component architecture, and an implementation plan. Those five documents live in the repo's docs folder and the build followed them. Writing the plan as files in the repo, not in a chat window, meant every later session could re-read the intent instead of relying on my memory of it.

2. 02

   Building the entire MVP in one stretch

   The first commit is the whole product: 109 files and roughly 19,700 lines covering the risk scan, career path recommendations, job matching, the 30/60/90 roadmap, resume and LinkedIn optimizers, the Kanban tracker, admin, and demo auth. I described features and reviewed what Claude produced; Claude wrote the code. It shipped with a deterministic mock AI engine behind a clean seam so the app was fully usable before any API key existed.

3. 03

   The 32-minute rebrand

   I decided FutureProof Jobs was the wrong name and asked Claude Code to rebrand to Career That Lasts. It renamed the product across the name, logo, page titles, docs, and user-facing deliverables in a single commit, 32 minutes after the initial commit. Doing a full rename by hand across a codebase that size is the kind of tedium that usually makes people keep a bad name.

4. 04

   Wiring real Claude behind the mock

   I asked Claude Code to replace the mock AI engine with real Claude API calls for the risk scan, resume optimizer, LinkedIn optimizer, and roadmap. It built server-only routes so the API key never reaches the browser, and used a forced tool call pattern so the model must return JSON matching an exact schema instead of free text I would have to parse. Every function keeps the heuristic engine as a fallback, so a failed API call degrades gracefully instead of erroring.

5. 05

   Real job listings, on the second try

   The first integration used the Remotive API for live job listings. Within half an hour of real use it was clearly wrong for this product: the listings skewed heavily tech, and this app is for everyone whose job AI might change. I had Claude swap the feed to Jobicy for cross-industry variety, then add an env-gated Adzuna source for local, all-industry listings, with honest labeling of remote-only results in the meantime.

6. 06

   Debugging match quality

   Job match scores looked off, so I dug in with Claude and found the skill extractor was only scanning the first 1,400 characters of each job description. Claude rewrote it to scan the full description and blended the fit scoring so matches stopped rewarding keyword stuffing near the top of a posting. This was found by eyeballing real results, not by tests.

7. 07

   Deploying and babysitting Vercel

   The app deploys to Vercel from the GitHub repo. The commit log shows the honest version of this: three separate empty commits exist purely to re-trigger Vercel deploys that did not pick up pushes on their own. Claude Code handled the git mechanics each time, but the deploy pipeline needed more nudging than any tutorial admits.

8. 08

   Turning billing off, on purpose

   Stripe Payment Links were wired into the pricing page early. Then I made the call to launch free: Claude put all paid features behind a NEXT_PUBLIC_BILLING_ENABLED flag, defaulting to off, so the whole app is free during early access and paid plans can be turned back on with one environment variable. Two days later I had Claude add a shareable scorecard for the free scan so results could spread on their own.

## stack

## what I learned

• →My first job-board API was the wrong one and I did not find out until real listings rendered. Remotive returned tech-heavy results for a product aimed at everyone, and I burned a morning swapping to Jobicy and adding an Adzuna option. Third-party data sources are the part of an AI app most likely to fail, and no amount of planning docs caught it.
• →Forcing Claude to respond through a tool call with a strict schema beats asking nicely for JSON. The API routes never have to parse free text, and the shape is guaranteed before the UI sees it.
• →Build the mock first. Shipping a deterministic fake AI engine behind the same interface as the real one meant the app worked end to end before I spent a cent on API calls, and it now doubles as the fallback when the real API hiccups.
• →Vercel deploys did not reliably trigger on push. The repo has three empty commits that exist only to kick the pipeline. Not a disaster, but it is friction nobody warns you about.
• →Rebrand early or never. Because Claude Code could rename the product across the entire codebase in one pass, changing the name 32 minutes in cost almost nothing. The same decision a month later would have been a project.