ankitclassicvision/ankit_shared_skills/agenttwin

AgentTwin, public version

AgentTwin is a lightweight diagnostic instrument for AI agent workflows. It produces a plain-English wellness report plus an operator-grade process map.

This public copy intentionally omits private rubrics, private examples, internal customer references, and organization-specific capture rules.

When to use

Use when the user is doing any of these on an AI agent, automation, or agentic workflow:

• Designing a new agent or automation
• Evaluating an existing agent or vendor proposal
• Modifying prompts, models, gates, or workflow components
• Diagnosing drift, hallucination, cost spike, missed escalation, or silent failure
• Accepting a build as ready for broader use
• Onboarding an engineer or operator to an agentic workflow

Do not use for generic AI questions, personal drafting, or pure code bugs unrelated to agent/workflow behavior.

Report shape

Produce a self-contained markdown or HTML report with two views:

1. Summary view

• Overall grade: A, B, C, D, or F
• Five property cards in plain English:
• Stays in lane
• Checks facts
• Checks before doing
• Nothing is hidden
• Has a stop button
• Current vs ideal flow
• Three-item action plan

1. Process map view

• Nodes with runtime type, owner, input, output, status, and failure mode
• Edges with trigger, contract, latency expectation, retry behavior, and audit trail
• Cross-cutting state, memory, observability, and override model
• Ranked recommendations by severity

Five closed-loop properties

• Bounded: the workflow has a clear box, allowed actions, refused actions, and stop condition.
• Grounded: important claims and decisions cite source evidence that can be rechecked.
• Gated: risky writes, sends, labels, money movement, compliance actions, and human-impacting outputs require approval or deterministic checks.
• Observed: runs leave useful traces, metrics, alerts, and replayable proof.
• Governed: an owner can pause, override, audit, improve, and retire the workflow.

Grading

• A: all five properties healthy.
• B: four healthy, one needs work, no broken property.
• C: at least one broken property or two needs-work properties.
• D: two or more broken properties.
• F: three or more broken properties, or any workflow can perform high-impact writes with no gate and no audit.

When uncertain between adjacent grades, downgrade.

Verification before delivery

Before delivering the report:

• Confirm the workflow boundary and topology are explicit.
• Mark unknowns rather than inventing facts.
• Scan the report for raw emails, phones, tokens, credentials, message bodies, raw rows, or unnecessary personal data.
• State whether any external side effects were performed. Default should be none.