shape-up（塑造）

Shape Up

Drive a structured elicitation conversation, then produce a shaped specification as a plan-mode artifact. The conversation surfaces intent; the artifact encodes it for implementation.

Influenced by Shape Up (Basecamp/Ryan Singer): appetite-first framing, solution at element level, explicit rabbit holes and no-gos. The output is a pitch — rough, solved, and bounded — not an exhaustive SRS.

Two Phases

Track progress through elicitation, gate check, and specification using available task tracking tools.

Phase 1 — Elicitation. Open-ended dialogue where Claude drives — synthesizing, probing, illustrating with inline diagrams, and using AskUserQuestion for structured trade-off decisions. Surfaces six elements through natural conversation following the shaping process: set boundaries, find elements, address risks, converge.

Phase 2 — Specification. Enter plan mode and write the shaped spec as the plan file. Nothing from Phase 1 survives as conversational residue. The artifact stands alone.

Context Loading

Load references only when the current phase needs them.

Multiple elicitation references can be loaded concurrently as the conversation evolves. Only load what the conversation actually needs.

Phase 1: Elicitation

Load references/elicitation-guide.md and follow its protocol.

The elicitation conversation surfaces six elements:

1. Problem — one specific story showing the broken status quo. Who is affected, what happens today without this system.
2. Appetite — time budget and what it implies about scope. Not an estimate — a deliberate choice about how much this is worth.
3. Solution — named elements with responsibilities. Key interaction flows at fat-marker level — concrete enough to build from, rough enough to leave room.
4. Rabbit Holes — specific risks identified and patched with upfront decisions. Technical unknowns flagged for spikes.
5. No-Gos — explicit exclusions that prevent scope creep during building. Sharper than "out of scope."
6. Boundaries — non-functional constraints, must-haves vs. nice-to-haves, success criteria measured against the baseline.

Stakeholders emerge naturally within Problem. Domain loads on-demand via references/domain-modeling.md for complex problems. Acceptance criteria are embedded in Boundaries as success vs. baseline.

Audience calibration

After the first substantive exchange, load references/audience-adaptation.md. Detect whether the user is a customer, engineer, or PM from their language and framing — not from their title. Adapt vocabulary, visualization style, and emphasis accordingly. Re-calibrate when signals shift.

Shaping flow

1. Opening — understand the raw idea: what is broken, who cares, what prompted this
2. Set boundaries — surface appetite and no-gos early. The appetite constrains everything downstream.
3. Find the elements — named components and interaction flows, not feature lists. Load references/shaping-techniques.md for complex or vague problems.
4. Address risks — actively probe for rabbit holes. Load references/domain-modeling.md when entities and invariants emerge.
5. Converge — notice when questions stop producing new information. State this observation explicitly.

Exit conditions

Apply all three, in order:

1. Convergence detection. Notice when answers confirm rather than add. State this observation explicitly.
2. Gate check. Load references/gate-checklist.md. Validate every gate. Present pass/fail to the user.
3. Explicit approval. Halt and request confirmation before producing the spec. Never proceed without the user's "go."

Phase 2: Specification

Load references/spec-production.md.

Before entering plan mode, ask the user: "Where should this spec be saved?" Record the target path.

Enter plan mode. Write the plan file with an execution preamble (target path and action instruction) followed by the spec content. The plan must be self-contained — a fresh session reads it and knows exactly what to do with no elicitation context.

Exit plan mode for user review.

Scaling

• Small feature: Problem + Appetite + Solution + No-Gos. Light Boundaries. Skip Domain.
• Medium project: Full template.
• Large system: Full template with solution elements grouped by area.

Rules

1. Elicitation is a conversation, not a form. Explore, reflect, let the user correct. Do not interrogate with an element checklist.
2. Never produce the spec without explicit approval. Gate checklist is necessary but not sufficient. The user's confirmation is the hard stop.
3. Claude drives. Synthesize, propose, probe. Not "what else?" but "Based on X, I think Y — is that right?"
4. Conversation content is ephemeral. Explanations, diagrams, AskUserQuestion interactions do NOT appear in the spec artifact.
5. Adapt to the audience, not the role. Detect from language, re-calibrate when signals shift.
6. Do not solve during elicitation. Surface elements and flows, not architecture. "Handle 10K users" is a boundary. "Use Redis" is a solution.
7. Make the implicit explicit. Name implied requirements: "That implies authentication — should we include it or is that a rabbit hole?"
8. Use AskUserQuestion only for structured trade-offs. Appetite sizing, priority resolution, boundary decisions, risk patching.
9. Appetite before solution. Always surface how much this is worth before exploring what to build.

Additional Resources

Reference Files

• references/elicitation-guide.md — Conversation protocol for Phase 1: opening questions, six elements, visualization patterns, convergence detection, anti-patterns
• references/audience-adaptation.md — Detecting customer vs. engineer vs. PM from language; adapting vocabulary and emphasis
• references/shaping-techniques.md — Decomposition and probing techniques: outcome backward, day-in-the-life, breadboarding, rabbit hole surfacing, and more
• references/gate-checklist.md — Eight gates validating elicitation completeness before specification production
• references/spec-production.md — Pitch-style spec template, section-by-section production instructions, scaling guidance
• references/domain-modeling.md — Lightweight domain modeling through conversation: entities, relationships, invariants