| name | grill-me |
| description | Interview Tobi relentlessly about a plan or design decision until reaching shared understanding, resolving each branch of the decision tree one fork at a time. Use when he wants to stress-test a plan before building, asks to "get grilled", or says "grill me". For design/architecture forks — not for mechanical edits. |
Grill the plan before any code
Claude's default failure mode is charging ahead on wrong assumptions. This skill flips it:
resolve the decision tree with Tobi first, then build. Use it before a non-trivial design
choice — the kind that would otherwise become an ADR.
How to grill
- Answer it yourself first. Before asking Tobi anything, check whether the docs or code
already decide it:
architecture.md (design wins, per CLAUDE.md), the ADRs in
docs/decisions/ (the decision log + phase→ADR map in progress.md point to the governing
one), progress.md itself, then the code. Only ask what genuinely needs his call.
- One fork at a time. Walk the decision tree branch by branch, resolving dependencies in
order — never dump a list. Each answer narrows the next question.
- Always recommend. For every question, give your recommended answer with a one-line why.
Use
AskUserQuestion (recommended option first, labelled "(Empfohlen)") for clean 2–4-way
forks; ask in chat for open-ended points. Questions in German; this file stays English.
- Be relentless, but converge. Keep going until the design is unambiguous — architecture,
data shape, edge cases, failure modes — but stop the moment there's nothing left that would
change what gets built. Don't manufacture questions.
Respect the golden rules while grilling
Surface tensions with the project's invariants as forks, don't silently assume them:
dice = code / narration = LLM (#2), the JSON-vs-recap memory split (#3), feedback protection
(#4), two-bot isolation (#5), system-agnostic-from-PDFs (#7). If a plan would cross one, that's
a question, not a given.
Close-out
When the tree is resolved, summarise the agreed design in a few lines and — if it settled a
real trade-off — offer to scaffold the next-numbered ADR (hand off to session-ritual). Then,
and only then, start building (or run /tdd if it's deterministic-core work).
