plan-interrogate
Drive a plan from sketch to commitment by resolving every open decision before any code is written.
Method
- Restate the plan in one paragraph. Confirm with the user that this is the plan being interrogated. Do not proceed on a mis-restatement.
- Extract the decision tree. Every branch point becomes a node. Mark
each node as open (undecided) or resolved. A resolved node
carries a source tag:
user(the user answered),inferred(the codebase or an existing constraint settled it). - Resolve in dependency order. A node is ready when every node it depends on is resolved.
- For each ready open node, ask exactly one question. Keep the question tight and binary or small-multiple-choice when possible.
- Pair every question with a recommended answer and one sentence of reasoning. The user can confirm, pick a different option, or push back.
- Before asking, check whether the answer already lives in the codebase,
prior commits, or an existing doc. If so, skip the question and mark
the node resolved with source
inferred: <path>. - Exit only when zero nodes are open. Print the resolved tree as a flat
list:
Decision - Choice - Source (user | inferred: <path>).
Anti-patterns
- Asking multiple questions at once. The user loses context and you lose the ability to react to each answer individually.
- Asking before exploring. If a fifteen-second read would answer the question, read first.
- Asking without a recommendation. A question without a stance is a survey; it offloads design onto the user.
- Rolling past an unresolved node. If a dependency is not pinned, the downstream question is premature.
Outputs
The interrogation produces three artifacts, not just answers. Offer to write each; do not force it.
-
Decision ledger (always). The resolved tree as a flat list:
Decision - Choice - Source (user | inferred: <path>). -
CONTEXT.md(when the interrogation surfaced project-specific terms). A short shared-language file: every domain term you and the user had to pin down, with a one-line definition in the project's own words. This is what stops the agent from using twenty words where one will do next session, and keeps names in code consistent. One term per line:term - what it means here. Point future sessions at it. On re-run, merge new terms in place rather than overwriting existing ones. -
Decision records (for contested or hard-to-reverse nodes only). One short record per decision that a future reader would question: the context, the choice, the alternatives rejected, and why. Keep them in
docs/decisions/NNNN-slug.md. Read the directory first and number from the highest existing record so two records never collide. Skip the obvious ones - a record for a trivial choice is noise.
Output contract
The decision ledger the user can paste into the plan doc. No prose summary.
No hedging. If the user declines to decide a node, mark it DEFERRED with
the reason the user gave - this is not the same as open. When you write
CONTEXT.md or a decision record, keep it in the project's language, not a
generic template.
