feat(browser): step byg-et-flow's intake into an ordered interview + don't build faster than you can test

[simo787c]

Two changes to byg-et-flow, both prompted by real bilagskontrol runs in the claude.ai browser. One file: agentbase-browser/skills/byg-et-flow/SKILL.md.

1. Stepped, ordered interview (replaces the flat five-point intake)

The old flat checklist made the model dump all ~5 question-clusters at once (a wall), ask deep edge-case questions in the abstract before seeing a sample, and once assert a block didn't exist without checking the catalog.

• Intake is now TRIN FOR TRIN in a fixed order: proces → sample tidligt → varianter → output → menneskeligt tjek → spørgerunde. Asking for a real example at step 2 grounds later questions in it instead of asking abstractly.
• New "Sådan spørger du" folds grill-me's principles in as prose (one question at a time, recommend an answer, explore-don't-ask, sharpen fuzzy terms, stress-test from the sample) — deliberately never via AskUserQuestion, which doesn't exist in the browser this pack targets.
• New catalog rule: never claim a block exists / doesn't without list_node_types/get_node_type; "menneske i løkken" means the flow delivers a result the human reviews (e.g. a control sheet with an empty status column), not a step that pauses the run.

RED/GREEN: RED (old skill) reproduced the multi-cluster dump + abstract edge-cases. GREEN interviewed one topic at a time in order, requested the sample early, ran a real spørgerunde with recommended answers, and handled the godkend-knap question without asserting node availability — prose only. Confirmed again on a live run: the spørgerunde caught a moms ekskl./inkl. trap that would have thrown a phantom 375 kr diff on every line.

2. "Byg ikke hurtigere end du kan teste" (tracer-bullet build discipline)

That same live run interviewed well but then built and saved the entire match-flow on a single faktura — no betalingsdokumentation, no Excel-oversigt — so the matching logic could never run, and it handed back a flow that was "gemt, men uprøvet". It also guessed an undocumented node config instead of probing it.

• New rule in the build loop: you need a sample you can run the flow helt igennem on before building the whole graph; with a partial sample, build and verify only the runnable slice now (e.g. OCR + sortering on the one bilag via run_node) and defer the rest; probe uncertain node config on real data with run_node instead of guessing; never save a full flow as "færdigt" when it has never run.

RED/GREEN on the exact "one faktura, rest comes later, fortsæt" decision point: a faithful (non-leading) RED with the interview-only skill reproduced the wild failure — it laid the whole 6-node graph in one set_flow and saved it as v0.1 untested, following "læg hele grafen i ét set_flow" literally, only then noticing it couldn't run. GREEN built only the OCR slice, ran it on the real faktura via run_node (surfacing that the filename key I2 differs from fakturanr 100447), and explicitly deferred the matching until a complete sag exists — saving nothing untested.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SYvwyQEa9xsKzsQTwpBuhp