/preflight

name: preflight
description: "Use before starting delivery work. Pre-implementation validation checklist to ensure readiness."
metadata:
  instruction_budget: "45"
  framework_dependency: "mycelium"
  framework_dependency_note: "This skill is designed to run within the Mycelium framework (https://github.com/haabe/mycelium). Standalone use will skip the canvas state, theory gates, and harness behavior the skill assumes. Install: /plugin install mycelium@haabe-mycelium."

Preflight Skill

Pre-delivery validation checklist. Run before every implementation task.

Checklist

Constraints (ALWAYS FIRST)

Before scoping any delivery work, establish constraints. Do not propose a plan before knowing the budget.

• [ ] **Time budget**: "How much time do you have for this?" (hours, days, sprint length)
• [ ] **Resource constraints**: Solo? Team? What skills are available?
• [ ] **Fidelity**: Demo/prototype vs. production? Who is the audience?
• [ ] **Dependencies**: Waiting on anything external?

If time budget < 8 hours, scope aggressively — one vertical slice, no polish. If the initial plan exceeds the time budget, cut scope before presenting it to the user.

**Re-forecast trigger (audit-triggered / emergent work).** Work that opens as "just address the recommendations", "quick fix", or any audit/assessment follow-up still gets a constraint pass — set an explicit estimate even when no one asked for one. Then, mid-session, **re-forecast when the work crosses ~2× the estimate or when no estimate was ever set**: stop, state actual-so-far vs estimate, and re-scope or re-confirm the budget before continuing. The failure this catches: emergent cycles that bypass preflight and balloon silently (dogfood `cycle-history.yml` — a "~2h" audit cycle ran ~9h; a "session-scope" one ran ~14h). The re-forecast becomes the `calibration.effort_accuracy` data point at `/retrospective`.

*Source: Hoskins transcript (2026-04-25) — agent proposed 20-hour plan before learning user had 8 hours. Goldratt (Theory of Constraints — identify the constraint before optimizing). Corrections.md: "Over-scope before constraints." Re-forecast trigger from the 2026-06-15 `/framework-health` effort-calibration finding (audit-triggered cycles balloon past estimate).*

Context

• [ ] corrections.md reviewed for relevant past mistakes
• [ ] patterns.md reviewed for applicable patterns
• [ ] Current diamond phase is Develop or Deliver
• [ ] Acceptance criteria are clear and measurable
• [ ] Scenarios linked: `.claude/canvas/scenarios.yml` has scenarios for this solution, and acceptance criteria trace back to scenario success/failure states (Hoskins)

Scope

• [ ] Scope fits within the time budget declared above
• [ ] This is the smallest vertical slice that delivers value
• [ ] Scope is explicitly bounded (what is NOT included)
• [ ] No speculative features (YAGNI check)
• [ ] Dependencies identified and available

Technical / Production Readiness

**Software/AI tool:**

• [ ] Tech stack detected and understood
• [ ] Build/test/lint commands confirmed working
• [ ] Development environment functional
• [ ] Existing code patterns reviewed

**Content:**

• [ ] Production tools ready (editor, recording, hosting)
• [ ] Style guide / editorial standards available
• [ ] Existing content patterns reviewed

**Service:**

• [ ] Delivery tools ready (templates, scheduling, communication)
• [ ] Existing service patterns reviewed

Security (software, ai_tool, service with digital infra)

• [ ] Data classification understood for this feature
• [ ] Security requirements identified (auth, input validation, etc.)
• [ ] No secrets will be hardcoded
• [ ] Dependencies checked for known vulnerabilities

Accessibility

**Software:**

• [ ] Accessibility requirements identified
• [ ] Semantic HTML approach planned
• [ ] Keyboard interaction model defined
• [ ] Color contrast requirements noted

**Content:**

• [ ] Captions/transcripts planned for audio/video
• [ ] Alt text planned for images
• [ ] Readable typography confirmed

Validation Strategy

**Software:** Test approach defined (unit, integration, e2e), edge cases and error scenarios identified. **Content:** Review process defined (SME, self-checklist, fact-check), learning objectives mapped. **AI tool:** Eval test cases defined, red-team scenarios planned, bias testing approach chosen. **Service:** Walkthrough planned, client feedback mechanism defined.

Success Criteria (G-V11)

• [ ] Declare what will be true after this delivery increment (1-3 measurable statements)
• [ ] Declare how each criterion will be verified (test, manual check, metric)
• [ ] Record criteria in .claude/harness/decision-log.md alongside the delivery decision

Example:

Success criteria:
1. "Users can complete onboarding in < 5 min" — verified by usability test
2. "API responds in < 200ms at p95" — verified by load test
3. "No new lint or type errors introduced" — verified by CI

Definition of Done

• [ ] DoD criteria reviewed and understood
• [ ] All criteria are achievable within this task

Before anything meets real people (v0.246.0)

A pilot is real people. Before the increment is deployed, published or sent to anyone, an exposure record on the diamond carrying it covers it, with **Security, Privacy and Service Quality passed** for that exposure (since v0.306.0 there is no fallback to the diamond's phase): `python3 "${CLAUDE_PLUGIN_ROOT}/scripts/scale_locks.py" --exposure-state`. If it prints what is missing, that is the next step, before any "pull and restart" instruction to the user. The exposure gate blocks the deploys the agent runs; this line is the guard on the ones the user runs.

**And record who it reaches** (v0.294.0): an entry in the carrying diamond's `exposures`, with the audience, channel, data class, end date, consent and the gates passed for it. `python3 "${CLAUDE_PLUGIN_ROOT}/scripts/scale_locks.py" --exposures` says what is missing. Report only until stage 2 of th