Importing A Codebase

Importing a codebase

The workspace holds real code but no specs. Reverse-engineer the spec graph the project should have had. When the repo already carries real spec-like documents, build the graph around them, not parallel to them. Do as much as possible yourself, from the files; ask the user only where the code genuinely can't tell you and the answer changes a spec.

Hold the writing-specs bar. Read that concept skill before drafting — everything in this flow is inferred rather than confirmed, so its honesty rules (draft until the user reviews, unconfirmed marked inline) bind hardest here.

1. Read first, ask last

Survey before you ask a single question. Read, in roughly this order:

• Agent files (mine these first — they state intent + conventions directly): AGENTS.md, CLAUDE.md, .cursor/rules/*, .cursorrules, .github/copilot-instructions.md, GEMINI.md, .windsurfrules.
• Docs: README, docs/, CONTRIBUTING, ADRs.
• Manifests & layout: package.json / pyproject.toml / go.mod / Cargo.toml, workspace globs, tree-style structure, entry points, build/test scripts.
• Code: entry points and the top of each candidate module — enough to see responsibilities and the dependency edges between them.

While you read, collect adoption candidates: durable, declarative documents that state the world as it is — architecture/design docs, ADRs / decision records, domain glossaries, protocol/contract docs. Never candidates (input only): READMEs, CONTRIBUTING, changelogs, roadmaps, TODOs, implementation plans (finished or planned), generated API docs.

Confirm with the spec tools (spec_grep / spec_graph) that there's no graph yet. If specs already exist, stop and hand back to the setting-up-a-project dispatcher — this flow is for un-specced repos.

2. Build a working model

From what you read, form a working model of what the project is and how it's shaped — held in the conversation, not written to a file (this flow declares no working files):

what:       one-sentence purpose (the job the codebase does)
domain:     the space it's in
stack:      languages / frameworks / runtime
modules:    the real boundaries + the dependency edges between them (who imports whom)
invariants: rules the code already enforces (layering, "X never imports Y", public surfaces)
decisions:  non-obvious choices visible in the code (and where the "why" is missing)

Agent files and READMEs usually hand you what, invariants, and decisions for free — prefer them over re-deriving from code.

3. Interview only the gaps

Ask only what the files can't answer and that would change a spec — typically: the primary job / who it's for, explicit non-goals, and the why behind a non-obvious decision. Interview them in rounds per the asking-user-questions concept skill; infer a concrete answer and let the user correct it rather than asking open-ended.

If adoption candidates exist, add one question to the same round: a multiSelect listing them (grouped when many — an adr/ set is one option) — which should become spec-graph nodes? A contradiction between a candidate and the code found by now goes into the round too (confirm the correction). Skipped or declined → adopt none; candidates stay input, noted at hand-off.

If the files answered everything material, skip the interview and say so — don't manufacture questions (adoption candidates alone still make a round — the offer is never dropped as "no gaps"). A skipped/declined question is not a blocker: record the assumption inline in the spec, marked unconfirmed.