Appearance

What This Skill Does ​

A rigorous, multi-section plan review from a CEO/founder perspective. Challenges premises, maps failure modes across 11 sections, traces error paths with severity tiers (BLOCKER/HIGH-LEVERAGE/POLISH), and ensures the plan is extraordinary -- not just adequate. Operates in four user-controlled modes: SCOPE EXPANSION (cathedral thinking), SELECTIVE EXPANSION (hold scope + cherry-pick), HOLD SCOPE (maximum rigor), and SCOPE REDUCTION (strip to essentials).

When to Use ​

  • User asks to "think bigger", "expand scope", "strategy review", or "rethink this"
  • A plan feels under-ambitious or is questioning its own scope
  • Before major implementation begins, to catch landmines early
  • When the user wants a second opinion on plan quality and completeness
  • NOT for idea validation before planning -- use mk:office-hours first
  • Use AFTER a plan exists

Core Capabilities ​

  1. Pre-Screen (Layer 0-1): Placeholder scan for unfinished sections (TBD, "implement later", "if time permits"), structural completeness check (5 required plan sections), and coverage mapping (every requirement mapped to a task). Never rejects outright -- returns for amendment with actionable guidance
  2. System Audit: Before the review, gathers full context: git history (30 commits), diff stats, stashed work, TODO/FIXME patterns, recently touched files, design docs, handoff notes, retrospective analysis, frontend scope detection, taste calibration, and landscape check (web search for industry context)
  3. Mode Selection: Four user-controlled modes:
    • SCOPE EXPANSION: Dream big -- 10x check (10x more value for 2x effort), platonic ideal, delight opportunities, expansion opt-in ceremony where user approves each proposal individually
    • SELECTIVE EXPANSION: Hold current scope as baseline, then cherry-pick ceremonies for each expansion opportunity
    • HOLD SCOPE: Complexity check, minimum changes assessment, maximum rigor with no scope changes
    • SCOPE REDUCTION: Ruthless cut to absolute minimum viable
  4. Implementation Alternatives (MANDATORY): Every plan gets 2-3 distinct approaches (at least: minimal viable + ideal architecture, optionally creative/lateral). User must approve before proceeding
  5. Two-Lens Evaluation: Lens A (Intent Alignment -- is this the right problem?) and Lens B (Execution Credibility -- can an engineer deliver?). Each grades PASS/WARN/FAIL independently. Any FAIL stops the review with NEEDS REVISION
  6. Deep Review (11 Sections): Architecture, Error & Rescue Map, Security & Threat Model, Data Flow & Interaction Edge Cases, Code Quality, Test Review, Performance, Observability & Debuggability, Deployment & Rollout, Long-Term Trajectory, Design & UX. Every section must surface at least one finding (adversarial necessity)
  7. Severity Tiers: Every finding classified as BLOCKER (must fix), HIGH-LEVERAGE (>20% outcome improvement), or POLISH (<5%)
  8. Outside Voice: Optional Claude subagent with fresh context for independent plan challenge -- catches structural blind spots
  9. Spec Review Loop: Adversarial subagent reviews the CEO plan document on 5 dimensions (Completeness, Consistency, Clarity, Scope, Feasibility) with up to 3 fix-and-redispatch iterations
  10. Verdict: Blockers > 0 -> NEEDS REVISION, else APPROVED with notes. Appended to plan.md (never overwrites). CEO plan persisted to ceo-plans/ directory for EXPANSION/SELECTIVE modes
  11. Review Readiness Dashboard: Shows all review statuses, staleness detection, and recommends next reviews

Arguments ​

Mode is selected interactively via AskUserQuestion (not CLI arguments). Context-dependent defaults: greenfield -> EXPANSION, enhancement -> SELECTIVE EXPANSION, bug fix -> HOLD SCOPE, >15 files touched -> suggest REDUCTION.

Workflow ​

Layer 0-1: Pre-Screen -> Step 0: Scope/Mode -> Step 0.5: Two-Lens -> Sections 1-11: Deep Review -> Outside Voice (optional) -> Layer 5: Verdict + Handoff

Detailed:

  1. Pre-Screen: Placeholder scan, structural completeness, coverage mapping. User decides: amend or proceed
  2. Initialize: Preamble, base branch detection, pre-review system audit (git history, design docs, handoff notes, TODOs, retrospective, frontend scope, taste calibration, landscape check)
  3. Prerequisite Skill Offer: If no design doc found, offer mk:office-hours inline. If user can't articulate the problem mid-session, re-offer
  4. Step 0 -- Nuclear Scope Challenge + Mode Selection: Premise challenge, existing code leverage, dream state mapping, 2-3 implementation alternatives (mandatory), mode-specific analysis with opt-in or cherry-pick ceremonies, temporal interrogation, mode selection
  5. Step 0.5 -- Two-Lens Evaluation: Intent Alignment + Execution Credibility. FAIL on either -> NEEDS REVISION, stop
  6. Sections 1-11 -- Deep Review: One section at a time. Each must surface >=1 finding. AskUserQuestion per issue -- never batch. Severity tiers on all findings
  7. Outside Voice: Optional adversarial subagent with fresh context. Cross-pass tension analysis
  8. Verdict: Blockers > 0 -> NEEDS REVISION. Append CEO Review block to plan.md
  9. Post-Review: Completion summary, review readiness dashboard, plan file review report, review chaining recommendations, docs/designs promotion (EXPANSION/SELECTIVE only)

Usage ​

/mk:plan-ceo-review

(plan file path provided in conversation context -- the skill detects it automatically)

Example Prompt ​

"I've written a plan for our auth system refactor but I'm worried it's under-ambitious. Can you think bigger?"

The skill will: pre-screen the plan for unfinished sections, run the full system audit, walk through the scope challenge with mode selection (likely EXPANSION for this request), produce 2-3 implementation alternatives, run the two-lens evaluation, then deep-review all 11 sections with severity-tagged findings.

Common Use Cases ​

  • Pre-implementation plan quality assurance
  • Scope expansion: "Is this ambitious enough?"
  • Architecture stress-testing before committing resources
  • Catching missing error paths, security gaps, and untested edge cases
  • Getting an independent, adversarial second opinion on a plan
  • Stripping over-engineered plans back to MVP essentials

Pro Tips ​

  • The pre-screen never blocks autonomously -- user always decides: amend or proceed with known gaps
  • Two-lens evaluation gates the deep review: if the plan has misaligned intent or isn't executable, the review stops early to save time
  • Every section requires at least one finding -- "no issues found" is valid only with documented evidence proving why the section is clean
  • The outside voice subagent has independent fresh context and catches blind spots the primary reviewer is structurally blind to
  • The review appends to plan.md -- it never overwrites. Both the original plan and the CEO review coexist
  • For EXPANSION/SELECTIVE modes, the CEO plan is persisted to .claude/memory/projects/ceo-plans/ and can be promoted to docs/designs/
  • The GATE 1 HARD STOP at the end means the agent does NOT chain into implementation automatically -- the human decides

Notes ​

  • The doc page is approximately 5% of the source content. Nearly every reference file's content is missing from the documentation.
  • The source has 13 reference files totalling ~15,000+ words of detailed process instructions. The doc has ~300 words.

Released under the MIT License.

Built with MeowKit