Founderealm Lens — Data skill for Claude Code

Founderealm Lens

Three files, one job each. Drop them into a repository they have never seen, run once, and they leave behind a small set of tools that answer questions about that repository with evidence attached. Then delete all three; what they planted keeps working without them.

file	what it is
founderealm_seed.py	the coordinator: survey the terrain, write the tools, ask before writing anything
founderealm_code.py	the product: every function that gets copied into the repository
lens_dna.json	the declarations: parser rules, directories to walk, tables and their caveats, dependency pins, the starter lenses

The split is the point. A change to the parser is a change to `founderealm_code.py` and lands in every tree. A change to what gets built is one diff in `lens_dna.json` and no Python at all. A change to how germination behaves stays in the seed. The filename tells you the blast radius before you open the file.

None of the three is read at runtime. Every generated file carries the declarations its own code names, so a germinated tree owes nothing to the thing that made it.

It reads source as text and bytes. It never imports, compiles, or executes the code it observes.

Built for AI coding agents. Claude Code, Cursor, Copilot, Aider and anything else working in a repository it did not write spend most of their context re-deriving the same structure: what exists, where it is defined, what calls it. This gives that agent a map to query instead, in SQL, and makes it answer with the limits of its own evidence attached. It works just as well for a person with a terminal; the agent is simply the reader that pays the highest price for a confident wrong answer.

Why

Working in a codebase means answering the same questions over and over. Where is this defined, what calls it, what breaks if it moves. Grep answers slowly and, worse, answers incompletely without saying so: a slice of the truth reads exactly like all of it, and