Projects from design documents
Give Claude a finished low-level design, and it builds the KPD project from it, with a report of everything it found.
KPD can be filled in without the UI. The lld-to-kpd skill, kept in the KPD repository at .claude/skills/lld-to-kpd/, tells Claude how to read a low-level design (LLD) PDF and any data-model or mapping workbooks, and build a project from them. It has been used on three real designs.
What you get back
- A project you can open, validate and export to Terraform, laid out in the document's own layers or steps, with notes on anything questionable.
- A gap report (
GAP-REPORT.md): what KPD couldn't express, the problems found in the document, every assumption made, and features KPD could add. - The builder script and its field tables in
_build/, so the project can be regenerated.
Using the skill
The skill builds projects with KPD’s own code, so it runs in Claude with a copy of the KPD repository rather than inside the beta. If you’d like a design document turned into a project, ask the KPD team. The result comes back as a project zip, which you bring into the beta with File → Import project (.zip).
With a copy of the repository, in Claude Code, ask for example:
Use the lld-to-kpd skill on ~/Downloads/Acme_Orders_LLD.pdf and the two mapping workbooks next to it.
What to give it
The more you include, the fewer assumptions the report lists:
- The design document (PDF). Diagrams are read too, and they often contradict the tables.
- Data-model and mapping workbooks: canonical model, source-to-target mapping, tagging matrices.
- What already exists: deployed schemas, existing topic names, sample messages.
- Standards the document refers to: naming, classification, access model.
Reviewing the result
- Open the project in KPD and check the canvas against the document's diagram.
- Read the Problems tab. Warnings in these builds are real findings.
- Open Export → Terraform for the permission advice and the TODOs the document didn't define.
- Treat the Flink SQL as a draft: it is generated from the mapping tables and hasn't been run.
- Work through the gap report's assumptions with the document's author.
The full playbook is in docs/LLD_TO_KPD.md in the repository.