ai-invert

/ai-invert — Dispatch the AI Inversion Analyst

Purpose

Entry point for AI-specific inversion analysis. Delegates the heavy lifting to the ai-invert-analyst agent on a separate thread so the main conversation stays lean.

Use this before implementation for any AI work that touches prompts, models, RAG pipelines, or AI-powered outputs. For non-AI code changes within an AI project, use /genai-invert instead.

Pre-flight: Auto-sync Check

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 If ## How to Approach Tasks appears AFTER ## Tech Stack or ## Architecture, the file has an outdated section order.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.

Input

The user provides a requirement, feature description, or task involving AI components.

Workflow

Step 1 — Confirm Scope with the User

Briefly restate the requirement and confirm the AI components touched (prompts, RAG, guardrails, outputs). If the requirement is a single sentence with no detail, ask the user whether they want /discover to elicit structure first — an inversion on a vague requirement produces vague risks.

Step 2 — Dispatch the Agent

Spawn ai-invert-analyst via the Agent tool (subagent_type: "north-starr-genai:ai-invert-analyst") on a separate thread. Pass the full requirement text plus any known constraints (cost envelope, latency target, deadline).

The agent will:

1. Read relevant code, prior decisions (.plans/DECISIONS.md), and learnings (.plans/LEARNINGS.md)
2. Work through 10 risk dimensions (user impact, prompt fragility, hallucination, data pipeline, cost, reasoning, model dependency, guardrails, architecture, virtue trade-offs)
3. Generate adversarial input examples tailored to this pipeline
4. Classify every risk NEW / PRE-EXISTING / AMPLIFIED
5. Cross-consult cost-estimator, rag-advisor, and guardrails-designer where the risk profile requires it
6. Write .plans/INVERT-<name>.md and return a summary

Step 3 — Read the Result and Present

When the agent returns, read .plans/INVERT-<name>.md and present a concise summary to the user:

AI inversion analysis complete: .plans/INVERT-<name>.md

Overall risk: <LOW/MEDIUM/HIGH>
Top 3 risks:
  - <risk 1>
  - <risk 2>
  - <risk 3>

Cross-consult log: <count> peer agents cited

Next step:
  - MEDIUM or HIGH → "Risk is <level>. I'll spawn `genai-layoutplan` to build an implementation plan. It runs on a separate thread so your main context stays clean. Proceed?"
  - LOW → "Risk is LOW. `genai-layoutplan` is optional. Want structured planning, or proceed directly to implementation?"

Step 4 — Dispatch genai-layoutplan (on user approval)

If the user approves and risk is MEDIUM or HIGH, spawn genai-layoutplan with the INVERT file path. It runs on a separate thread.

Notes

• This skill is a thin dispatcher — the real analysis lives in the ai-invert-analyst agent
• The agent produces .plans/INVERT-<name>.md which feeds genai-layoutplan
• For non-AI code changes within an AI project, use /genai-invert (standard inversion)
• The agent runs on a separate thread to keep main conversation context clean
• If .plans/INVERT-<name>.md already exists for this requirement, tell the user and ask whether to refine (add a revision section) or regenerate