Why there is an interview
The spec route is only as good as what it is given. A process flow captures step order, decisions and failure paths, and a sequence diagram captures the calls, but a build spec also needs the stack and the exact commands, actors with permissions per entity, typed fields, edge cases, external systems, numeric targets, a sign-in method and a design direction. That knowledge lives with you, so Runic asks before it writes. The interview is stage one of the Build Package; the spec is stage two, and it receives your answers together with the documents, the brief, the plan and the ID INDEX of every branch and message.
What it asks
After you click Start, the ticked processes and sequence diagrams go to the model with Runic's reading legends, so captions, .fail outcomes, .back arrows and .alt sections are read as what they are. The prompt casts the model as a senior business analyst preparing a build specification and asks for five to eight questions, each specific to these documents: it must quote the step or process it is asking about, exactly as written. The gaps it is told to cover: actors and roles with who may take each decision and who may see what, the data each step reads or writes, failure and edge cases, external systems, and volume or constraints. Then the things the package itself needs, asked only when neither the brief nor the flows answer them: the stack (language, framework, database) and the exact commands to install, run, test, lint and typecheck; the numbers behind the non-functional requirements (p95 latency of the core call, monthly availability, data retention, session length, concurrent users); how users sign in (password, OTP, SSO or magic link) and which field identifies them, because Claude Code cannot start the first milestone without it; the target platform and three to five words for the tone of the design. A stack, number or design question has to say what the package will do with the answer.
Two fidelity rules shape the questions. Names are quoted character for character, never paraphrased or translated, and every number, threshold, role and system the brief or a flow states is kept verbatim. When you arrived from Whole system, the brief is what you originally asked for and the flows were drafted from it: anything the brief states that the flows do not carry yet is worth a question, anything either already answers is never asked again, and a plan assumption is asked exactly once, as a yes or no confirmation quoting it, and only when the answer would change a diagram. A document that was cut is marked with “[… N more lines not shown …]” and the model is told to treat that part as unknown, never as absent.
Questions come back one per line. Runic strips any numbering or bullets the model added and keeps at most eight. The prompt asks for them in the language the documents are written in.
How answers reach the spec
Each question gets a two-line answer box you can drag taller. When you click Generate package, every question is sent with its answer; a blank one is sent as (not answered). The spec prompt has a rule for that case: anything still unknown goes into open_questions instead of being invented. That list renders as the Open questions section of docs/spec.md, and both CLAUDE.md and KICKOFF.md make it one of the four reasons Claude Code may stop and ask you. The spec stage reads at most 12 question and answer pairs, each question cut at 500 characters and each answer at 2,000; the full text still lands in the appendix.
A stack answer becomes the Stack & commands section of CLAUDE.md and, when it names a test command, the .claude/settings.json hooks. A roles answer becomes the roles by permissions matrix. A sign-in answer becomes the spec's auth block, the Sign-in line under the matrix, a convention line in CLAUDE.md, and it removes the “Before M1: the package does not say how users sign in” ask from the kickoff. A number in a non-functional answer becomes an NF row with a comparator and a unit; an adjective becomes a failed quality tell. Tone and platform answers land in sections 3 and 4 of the design brief. The full interview is appended to the spec verbatim under Appendix: Interview, every question and answer as a blockquote, so the reasoning behind the spec travels with it.
What the model sees
Before the documents, when they exist: a BRIEF block with your original text verbatim (up to 12,000 characters), PLAN ACTORS with each actor's kind and goals, PLAN ASSUMPTIONS marked as unconfirmed, PLAN ENTITIES with each entity's complete status vocabulary, and PLAN RULES, the facts each document must carry. Then PROCESS FLOWS and SEQUENCE DIAGRAMS, each under its name. Every document is cut at a line boundary at 20,000 characters, even when it is the only one; when the set still outgrows 120,000 characters the longest are cut further, and every cut is announced. Nothing outside the ticked documents, the brief and the plan is sent: not your other documents, not the chat, not the memory.
Limits and gotchas
The questions stage runs on the mid tier of your provider (claude-sonnet-5 on Anthropic) with extended thinking, 1,500 output tokens and 120 seconds at most; the spec stage on the heavy tier, which defaults to claude-sonnet-5 too unless the operator sets ANTHROPIC_MODEL_HEAVY. Both count against one limit of 12 calls per hour per account and, on Runic's key, against the account's $5-per-month AI budget, which once used up answers 402 “This month’s AI budget on Runic’s key is used up — add your own key under Settings → AI or wait for the 1st.” More than 40 processes or 20 sequence diagrams is refused with a message naming the caps, not silently cut. There is no re-ask button: close the window and click Build again for a fresh set. This is a different interview from the questions that appear in the chat after Whole system returns a plan; those are bound to a document and turn into edits.
How to use it
- Pick the documentsClick Build in the top bar and tick the processes and sequence diagrams to include. With nothing ticked, Start refuses with Tick at least one process, the spec is built from what you select.
- StartClick Start. The window shows Reading your flows… and Finding what a developer would still ask. Esc and the ✕ are inert until the questions arrive.
- Read the questionsThey arrive numbered, each with its own answer box. Drag a box's corner to make it taller.
- Answer what you canThe note above the list says it: unanswered questions land in the spec's Open Questions so nothing gets invented.
- GenerateClick Generate package. If the AI call fails, the error shows in an amber note and you land back on the answers with everything you typed intact.
- Check the tellsOn the done screen, the Package quality panel shows which of the ten checks the answers satisfied. A failed Numbers in NFRs or Role closure tell points back at an interview answer.
Commands & shortcuts
| Start | Reads the ticked documents, the brief and the plan and returns up to eight questions. |
| Generate package | Sends documents, answers and the ID INDEX to the spec stage. |
| Esc | Closes the window between stages; ignored while a request is in flight. |
| Open questions | The docs/spec.md section holding what you skipped, what the model could not settle, and the NF gap questions. |
| Appendix: Interview | The docs/spec.md section with every question and answer verbatim, blockquoted. |
Tips
- Answer with names and numbers, not adjectives. Finance lead approves refunds over 500 gives the spec a role, a permission and a threshold; the right person approves gives it nothing.
- Give the stack as commands: npm install, npm run dev, npx vitest run, npx vitest run -t, npm run lint, npx tsc --noEmit. Each one lands in CLAUDE.md verbatim, and the test command wires the hooks.
- Say how users sign in, even in four words. Without it the spec carries an NF gap question and the kickoff makes Claude Code stop before Milestone 1 to ask you.
- Skip honestly. A blank answer is folded into Open questions, which the kickoff names as one of the four reasons Claude Code may stop and ask.
- Long answers are fine, but the model reads the first 2,000 characters of each and at most twelve pairs. The full text still lands in the spec's appendix.
- If the questions feel generic, the documents are thin. Add the decision, the .fail path or the .alt section in the editor, close the window and click Build again.
Questions
Can I regenerate the questions?
Close the window with Esc or ✕ between stages and click Build again. Each Start is a fresh read of the ticked documents; earlier answers are not kept.
Is this the same as the questions in the chat after Whole system?
No. Those come from the plan, are bound to one document and are answered in the chat so they change that document. The build interview feeds the spec and never changes a document.
What does the model see?
The text of the ticked processes and sequence diagrams, up to 40 and 20 of them, 20,000 characters each and 120,000 in total, plus Runic's reading legends and, when you arrived from Whole system, the brief and the plan's actors, assumptions, entities and rules.
How many questions?
The prompt asks for five to eight; Runic keeps at most eight.
Why does it ask about sign-in?
Because the first milestone cannot be planned without it. The answer fills the spec's auth block; without one the spec lists an NF gap for sign-in and the kickoff tells Claude Code to ask you for the method and the identity field before M1.