genai-invert

name	genai-invert
description	Deep structured inversion analysis on a requirement before implementation. Identifies risks, edge cases, failure modes, and convention violations. Use before complex or high-stakes tasks.
argument-hint	<requirement or feature description>

Inversion Analysis — What Could Go Wrong?

Purpose

Before committing to implementation, systematically invert the requirement: instead of "how do I build this," ask "how could this fail?" Produces a structured risk analysis that feeds directly into planning.

Run this before implementation for complex or high-stakes work. For routine tasks, the lightweight risk assessment in the project CLAUDE.md is sufficient.

Pre-flight: Auto-sync Check

Before starting, read the root CLAUDE.md (if it exists). If ## How to Approach Tasks appears AFTER ## Tech Stack or ## Architecture, the file has an outdated section order. Silently run /genai-sync to fix it before continuing with this skill.

Input

The user provides a requirement, feature description, or task. Can be a single sentence or a detailed spec.

Workflow

Step 1: Understand the Requirement

Actions:

1. Restate the requirement in your own words — confirm understanding
2. Read relevant code to understand what exists today
3. Check root context files (CLAUDE.md, AGENTS.md) for architecture, grain, and module map
4. Identify which modules and layers this task touches
5. Surface assumptions — what does this requirement take for granted? List every assumption the requirement makes about:
   • Input data (format, quality, availability, freshness)
   • User behavior (technical skill, workflow, frequency of use)
   • External systems (uptime, API stability, version compatibility)
   • Domain rules (regulatory stability, business logic, compliance criteria)
   • Infrastructure (latency, storage, compute budget) For each assumption, ask: "What happens if this is wrong?" Assumptions that would cause HIGH impact if wrong become risks in Step 2.

Step 2: Inversion Analysis

Systematically work through each dimension:

A. User / Consumer Impact

• What could frustrate or confuse the end user?
• What existing workflows could this break?
• What happens if the user does something unexpected?
• What accessibility or performance impact could users feel?
• Who is most harmed if this feature produces wrong output — and how badly? (e.g., financial loss, legal liability, regulatory penalty, reputational damage, safety risk). Name the specific stakeholder roles.
• Are there domain-specific consequences beyond software failure? (e.g., regulatory non-compliance, contractual breach, audit failure, licence revocation). If yes, these risks are automatically HIGH severity regardless of technical likelihood.

B. Technical Failure Modes

• What could crash, hang, or corrupt data?
• What race conditions or concurrency issues are possible?
• What happens under load, with bad input, or with no network?
• What external dependencies could fail? (API rate limits, model version changes, service outages, vendor lock-in)

If this feature uses AI/ML (LLMs, RAG, embeddings, classifiers, or any model inference):

• Hallucination: Could the model generate confident but factually wrong output? What are the consequences if it does?
• Stale retrieval: Could the RAG pipeline return outdated or irrelevant context? What happens if the knowledge base is out of date?
• Prompt injection: Could user-provided input manipulate the model to bypass business rules, leak data, or produce unintended output?
• Model drift: If the underlying model is updated or swapped, could output quality degrade silently? Is there a baseline to compare against?
• Confidence calibration: Does the system surface uncertainty to users, or does it present all output with equal confidence? What happens at low-confidence thresholds?
• Data pipeline integrity: At each step of the pipeline (ingestion → parsing → embedding → retrieval → generation → post-processing), what could go wrong? What validation exists between steps?

C. Data Flow & Pipeline Risks

• Map the data flow — trace data from entry to final output. What are the steps? (e.g., upload → validate → transform → store → notify, or ingest → parse → embed → retrieve → generate → review)
• At each handoff between steps: what could be lost, corrupted, or delayed?
• What happens if a middle step fails — does the pipeline retry, roll back, or leave partial state?
• Is there an idempotency guarantee if the same input is processed twice?
• What validation exists between steps — and what's missing?

D. Edge Cases & Boundary Conditions

• What boundary conditions exist? (empty, null, max, overflow, unicode, etc.)
• What state combinations are unusual but possible?
• What happens on first run, last run, or after a long gap?
• What platform or environment differences could matter?

E. Architecture & Convention Risks

• Does this go against the grain? (check grain section in root context files)
• Does this violate any path-scoped rules (.claude/rules/, .github/instructions/)?
• Does this create coupling between modules that were independent?
• Does this introduce a pattern inconsistent with existing code?

F. Virtue Trade-offs (Code Quality)

Check this change against the 7 Code Virtues (see skills/_references/virtues/code-virtues.md) in priority order:

• Working: Could this break existing behavior? Are there tests proving current correctness?
• Unique: Does this duplicate logic that already exists elsewhere?
• Simple: Does this add unnecessary entities, relationships, or indirection?
• Clear: Will the result be obvious to the next reader, or puzzling?
• Easy: Will this make future changes harder or easier?
• Developed: Are the abstractions mature and well-placed, or primitive?
• Brief: Is there unnecessary verbosity?

When two virtues conflict, always preserve the higher-priority one. Name the tension explicitly in the risk output:

**Virtue Tension:** [lower virtue] would improve, but at the cost of [higher virtue] → preserve [higher virtue]

G. Observability & Recovery

• If this fails in production, how would anyone know?
• Can the failure be diagnosed from logs/errors alone?
• Is the change reversible? What's the rollback path?
• What monitoring or alerting should exist?

Step 3: Assess and Prioritize

Rate the overall risk:

Level	Meaning	Action
LOW	Well-understood, contained, reversible	Proceed to implementation
MEDIUM	Some unknowns, but manageable with care	Plan carefully, validate incrementally
HIGH	Significant unknowns, wide blast radius, or irreversible	Spawn the genai-layoutplan agent to break into tracked pieces, spike first, or clarify requirements

Step 4: Produce Output

Present the analysis:

## Inversion Analysis: [requirement summary]

**Modules Affected:** [list]
**Against the Grain?** [yes/no — why]
**Virtue Tensions:** [any virtue trade-offs identified, or "none"]

### Risks

1. **[risk name]** — [severity: HIGH/MED/LOW]
   [description, how it could happen, what the impact would be]

2. **[risk name]** — [severity: HIGH/MED/LOW]
   [description, how it could happen, what the impact would be]

[...repeat for each significant risk]

### Assumptions (verify before implementing)

- [assumption]: [what happens if wrong]
- [assumption]: [what happens if wrong]

### Edge Cases to Handle

- [case]: [what should happen]
- [case]: [what should happen]

### Recommendations

**Overall Risk:** [LOW / MEDIUM / HIGH]

**Before implementing:**
- [prerequisite or clarification needed]

**During implementation:**
- [specific thing to watch for]
- [specific validation to include]

**Test strategy:**
- [specific test approach for each HIGH/MED risk — e.g., golden file tests, integration tests against real pipeline, adversarial input tests, human review sampling, load tests, A/B comparison against manual output]
- [what constitutes "passing" — acceptance threshold, not just "it works"]

**After implementing:**
- [what to verify]
- [what to monitor — specific metrics, dashboards, alerts]

Step 5: Persist to Disk

After presenting the analysis to the user (and incorporating any feedback), write it to disk so downstream agents can consume it without context loss.

Actions:

1. Create .plans/ directory if it doesn't exist
2. Generate a short kebab-case name from the requirement (e.g., auth-refactor, api-pagination)
3. Write the full analysis to .plans/INVERT-<name>.md using this format:

# Inversion Analysis: <requirement summary>

**Created:** <date>
**Overall Risk:** <LOW / MEDIUM / HIGH>
**Modules Affected:** <list>
**Against the Grain?** <yes/no — why>
**Virtue Tensions:** <any virtue trade-offs identified, or "none">

## Risks

1. **<risk name>** — [severity]
   <description>

[...all risks]

## Assumptions (verify before implementing)

- <assumption>: <what happens if wrong>

## Edge Cases to Handle

- <case>: <what should happen>

## Recommendations

**Before implementing:**
- <items>

**During implementation:**
- <items>

**Test strategy:**
- <specific test approaches for each HIGH/MED risk>

**After implementing:**
- <items>

4. Inform the user: "Analysis saved to .plans/INVERT-<name>.md"

Step 6: Trigger Planning

If the overall risk is MEDIUM or HIGH, prompt the user:

"Risk is [MEDIUM/HIGH]. I'll spawn the genai-layoutplan agent to build an implementation plan from this analysis. It runs on a separate thread so your main context stays clean."

Then spawn the genai-layoutplan agent (available in .claude/agents/ or .github/agents/). The agent runs on a separate thread to keep the main context clean.

For LOW risk, inform the user that the genai-layoutplan agent is available if they want structured planning, but it's optional.

Notes

• This skill is language-agnostic — it works for any project type
• Read actual code before forming opinions — never invert based on assumptions
• Root context files (CLAUDE.md, AGENTS.md) and path-scoped rules (.claude/rules/, .github/instructions/) provide the baseline for convention checks
• Focus on risks that are likely and impactful — don't enumerate every theoretical failure
• If the analysis reveals HIGH risk, recommend spawning the genai-layoutplan agent to break the task into tracked, safer pieces rather than proceeding with the full scope
• The output of this analysis feeds directly into the genai-layoutplan agent — the risks become constraints and dedicated tasks in the implementation plan