Job Synthesis
A local platform that turns one master profile and a job posting into an honest, tailored application: fit check, resume, cover letter and tracking. Claude, Codex and Antigravity all run the same workflow on it.
Who it’s forOne user so far: me. A test workspace of sample candidates stands in for demos.
ProductBuilt with AI
4days from first commit to a working platform
3AI agents on one workflow: Claude, Codex, Antigravity
6job boards it reads: Apple, Greenhouse, Indeed, Lever, LinkedIn, Workday
0databases: the files are the source of truth
The problem
- Every application needs a tailored resume and cover letter, and tracking dozens by hand breaks down.
- AI tailoring is fast, but it drifts: inflated titles, invented skills.
- I wanted different models to do the work and be compared, without one overwriting another.
What I did
- Wrote the synthesis workflow as an agent skill: a six-step path from master profile to fit check, resume, cover letter and pitch page.
- Added importers for six job boards that turn a posting into structured text and metadata.
- Wrote security regression tests: hostile headers, path traversal, shell injection and SSRF.
- Kept a living backlog of the next product upgrades.
What changed
- Every application gets a fit check before any writing starts.
- Every tailored package traces back to the master profile.
- A product I use myself, with a test workspace of sample candidates for demos.
Decisions
- The files are the source of truth.No database. The dashboard, the scripts and three different AI agents all read and write the same folders, so no tool owns the data.
- The fit check comes before any writing.The posting is scored against the profile, by tier, before a resume or a letter is drafted.
- Honesty is checked, not hoped for.A ledger lists the adjacent skills I have confirmed, and an audit catches job-title drift in the text a tracking system reads.
- A second model is benchmarked apart.Its output is compared in isolation and can never overwrite the baseline.
- Delete hides; it does not destroy.An application removed by mistake can come back.
How AI was used
- Three agents from three vendors run one workflow, written once as a skill, so I can compare them on the same job.
- The repository is the contract between them: each agent reads and writes the same profile and application folders.

