feat: add pre-flight check, customer discovery phase, and staged financials

[llillo]

What this PR does

Addresses three structural weaknesses in the startup-design skill for founders who want to validate an idea before investing heavily in brand, product, and financial planning.

Changes

1. Phase 0.5 — Pre-Flight Check (new)
A fast 5-minute sanity check before intake. Runs 3 searches to surface: dominant incumbents, prior failures of similar ideas, and obvious regulatory blockers. Gives the founder critical context before the process starts. Always runs, even in Fast Track.

2. Phase 3.7 — Customer Discovery (new)
Inserts real customer interviews between the Research Gate and Strategy — exactly where they belong. Includes an Interview Gate (analogous to the Research Gate) that pauses the process if fewer than 3/5 interviewees confirm the problem. Prevents investing in brand, product, and financials before validating the problem is real.

3. Phase 7 — Two-stage financial model (updated)
Splits projections into Stage A (Assumption-Based, default) and Stage B (Evidence-Based, when interviews produce willingness-to-pay signals). Every financial file now declares its stage explicitly, so the founder always knows how much to trust the numbers.

4. New reference file: customer-interview.md
Full interview protocol: who to interview, how to find them, 25-30 min question structure, notes template, cross-interview synthesis guide, and go/no-go criteria. Based on Rob Fitzpatrick's Mom Test principles.

5. Updated output-guidelines.md
Final Assessment Dashboard now includes two new fields: Customer Interviews (count + confirmation rate) and Financial Stage (A or B).

Why these changes

The original skill is excellent at designing a startup but optimized for thoroughness over speed-to-validation. A founder with a raw idea risks spending hours on brand identity and detailed financial projections before ever speaking to a potential customer. These changes insert two explicit validation gates (research + interviews) before the expensive phases, and make the financial model honest about its reliability.

Files changed

• startup-design/SKILL.md — updated flow, new phases 0.5 and 3.7, updated phase 7
• startup-design/references/output-guidelines.md — updated dashboard
• startup-design/references/customer-interview.md — new file
• .claude-plugin/marketplace.json — version bump to 1.8.0

[ferdinandobons]

Merged — thank you, this is a great contribution. The validation-first direction (pre-flight check, customer discovery gate, staged financials) fits the skill's radical-honesty philosophy perfectly.

A few follow-up adjustments landed on main right after the merge (4a60fbe, a61e2bd):

• Phase counts aligned — 8 numbered phases plus pre-flight and customer discovery; the Full Mode line and the PROGRESS.md checklist were still saying "8 phases" while the intro said 9.
• Phase 0.5 ordering — the pre-flight runs before the project directory and PROGRESS.md exist, so it can't save files yet. Findings are now presented in conversation and persisted during the Phase 1 output step.
• Interview Gate threshold gap closed — "4+/5 proceed" vs "fewer than 3 reassess" left exactly 3/5 undefined. Now: 4+/5 proceed, 3 or fewer reassess.
• Fast Track minimums aligned — SKILL.md said "at least 3 interviews", customer-interview.md said "treat as 0 and run the full protocol". Now consistent: minimum 5 in Full Mode, minimum 3 in Fast Track.
• Interview gate made explicitly deferrable — the founder can choose to continue on research data alone, with consequences stated plainly (financials stay Stage A, dashboard shows "0 conducted · deferred", interviews become experiment Bug: AGENTS.md not equal to CLAUDE.md #1 in the validation playbook). This keeps the skill's "one afternoon" promise intact while preserving the honesty pressure your PR introduced.
• "Field definitions" header restored in output-guidelines.md.
• The version bump was dropped from the merge (1.8.0 had already shipped in May); your changes will go out with the next release.

SKILL.md was also brought back under the 500-line limit after the additions (intake questions and phase output specs moved to references/). Thanks again!