Design interview

/design-interview — elicit design, then principles

Read ${CLAUDE_PLUGIN_ROOT}/skills/shared/tracking-rules.md first and obey it (especially: question gates, contextual chip construction, the DESIGN.md ownership boundary, and IP/GP numbering). Phase header: # Design interview → ## Facts / ## Principles. Chapter markers: mark a chapter at each phase transition (session start implicit).

This skill does what code-reading cannot: it interviews the human for the intent behind the repo and writes it into cairn/DESIGN.md. It runs in two phases with an explicit seam. Facts before principles is a one-way dependency — commitments made before the fact base exists get reworked once it does — so Phase 1 never asks the user to commit a principle; it only banks proto-principles for Phase 2.

Run this on Fable. The interview is markedly better on Fable than on Opus — its questions land in plain language and its Phase-2 judgment is stronger (openac pilot, 2026-07-11; D-014). If the session is not already on Fable, recommend switching before starting; then proceed. This is the user's per-instance model choice (no cairn-spawned Fable subagent, so D-004 is unaffected), not a standing elevation.

Session start

Require cairn/DESIGN.md to exist; if there is no cairn/, stop and route to /cairn-init (scaffold first, interview second). Read the current DESIGN.md, cairn/DECISIONS.md, and — so every option can be grounded in evidence — the repo itself: DESCRIPTION/manifest, the exported surface, Imports/dependencies, installer/platform coverage, and recent git log. For a repo spanning many files, fan out [S]-tagged Explore subagents with specific focuses and require file:line citations. Draft your grounded option sets internally before asking anything.

Also check for a migration-preserved numbered-principles file — a note-and-leave doc at a cairn/ path (e.g. cairn/PRINCIPLES.md) kept with numbering and basename intact because package code cites it by number; the migration ledger and the DESIGN seed point to it. When present, announce that this run will ingest it — see "Ingesting a note-and-leave principles file" below.

Phase 1 — Facts

Elicit what cannot be inferred from the repo; do not classify anything. Run a series of batched AskUserQuestion rounds (2–5 questions each, every option carrying a marked recommendation). Five disciplines, all mandatory:

1. Elicit, don't classify. Ask for intent the code can't show: who the package is for; where its job ends (the contract boundary); which platforms are commitments vs. best-effort; what bar a new capability must clear to earn a place; distribution ambition; API-stability posture; dependency posture.
2. Chain rounds on prior answers. Each round consumes the last — e.g. "You said the roster grows opportunistically; what should the bar be?" Batch only what is mutually independent within a round.
3. Ground every option in repo evidence. Cite real files, functions, Imports, installer coverage, recent commits in the option text (e.g. "os_fix_csv is the seed of this family").
4. Options are hypotheses that do work. Each option states an implication or consequence ("this sets the API-stability bar"), never a generic label.
5. Ask the wart question. "What warts or fragilities do you know about that code-reading wouldn't reveal?" — with options that are evidence-based guesses (recent fixes, unmaintained upstreams, platform-biased testing).

Bank, don't decide. When an answer implies a principle, name it and add it to a running banked-candidates ledger ("sounds like a principle — banking it for the principles round"); never ask for the commitment now.

The seam

At the phase boundary: summarize what was heard; write the DESIGN draft (Purpose & Scope, contract boundary, conventions) plus the banked-candidates ledger to disk; checkpoint-commit (stop points are commit points, so a later session resumes statelessly); then a continue/stop decision chip —

• Continue into principles → Phase 2 (recommended)
• Pause here — stop; resume statelessly later

Continuous by default, pausable without loss. Both phases run in one session — best on Fable (see the note at the top).

Phase 2 — Principles

Formalize and pressure-test. Sources for candidates: the banked ledger, plus two the interview must add itself —

• Mine git history for implicit principles. Generalize ad-hoc fixes into candidates (a "skip files without audio" fix → "resilient batches").
• Derive candidates from the domain, not just the code. State forward consequences (irreplaceable recordings → inputs-sacrosanct; IRB/consent → local-only processing, which constrains future remote-inference wrappers).

Then propose and test:

• Every candidate arrives classified. Proposed strength — IP (inviolable) / GP (guiding) / skip — with a marked recommendation, and adjacent-but-separate matters fenced off explicitly ("output-overwrite defaults are a separate, tradeable matter") so each question stays decidable.
• Stress-test the adopted set against phase-1 decisions. Surface collisions between separately-endorsed commitments ("thin wrappers" vs. a "tidy outputs" choice → where's the line?).
• Separate essence from accident. Ask whether a principle is the capability or the current idiom, citing recorded decisions as leverage.
• Probe the scope of each IP. Offer extend / keep / downgrade for every inviolable.

Write-out. Record the settled principles in cairn/DESIGN.md under Design Principles: the IP block first, then GPs; numbers run within each type and are never reused or renumbered (retiring one takes a D-entry). A genuine cross-cutting choice with rationale (not a deferral) becomes a cairn/DECISIONS.md D-entry. Deferred candidates become ROADMAP candidate rows — nothing the user surfaced is silently dropped.

Ingesting a note-and-leave principles file

A Lineage A migration may have preserved a numbered-principles doc at a cairn/ path with numbering and basename intact, because in-code citations (PRINCIPLES.md #N — 70× across 29 files in the M41 intraclass pilot) make folding it into DESIGN's IP/GP a renumbering that strands every ref (the migration protocol's forced note-and-leave). When session start found one, run these additions on top of the normal two phases:

• Every numbered principle enters Phase 2 as a candidate. Each arrives classified like any other candidate (proposed IP / GP / skip with a marked recommendation) and carries its #N lineage in the option text, so the user always sees which preserved principle is being formalized. The banked ledger, history mining, and domain derivation still run — ingestion adds candidates, it never replaces the other sources.
• Conservation: no ingested principle is silently dropped. Every #N ends the interview with an explicit disposition — formalized as an IP or GP, merged into another candidate, or retired at the user's say-so (IP3). A #N with no disposition is an unfinished interview, not a skip.
• Write-out records the lineage map. Under Design Principles in the target cairn/DESIGN.md, record an old-#N → new-id mapping table (#N → IPn | GPn | retired | merged into …) alongside the principles, so every in-code citation has a written bridge to its formalized successor.
• The preserved file stays intact. Its numbering and basename are never edited, renumbered, or deleted by this interview (IP4) — in-code citations must keep resolving against it until the repoint ships.
• Bank the repoint; never touch code. The eventual in-code repoint (PRINCIPLES.md #N → DESIGN.md IPn/GPn) is a target-repo code milestone: bank it as a candidate row in that repo's ROADMAP at write-out. The interview itself performs no code edits.

Routing

Close with the close block (tracking-rules "Question gates and phase closes") composed from what was produced — recap, status line, fenced next command(s) with plain labels (e.g. /milestone-plan to plan the first milestone, /milestone for a health check), and the adjust-or-/clear safety line; no chip.