Digital twin design planner

What it does

Bring a real system and its data sources. Set freshness and proof rules, then get a twin plan with action limits.

What goes in and out

You give it

• The system or process the twin will copy.
• The places its data comes from.
• The decision the twin should help with.
• How fresh and accurate the data must be.
• The owner, the risk level and the actions the twin may take (optional).
• Source text (optional)

You get

• A label for your idea: model, shadow or twin.
• A plan for identities, data formats, timing and proof of origin.
• A test plan that checks the twin against real results.
• A checklist of proof needed before the twin may act.
• A list of open gaps with owners.

How the task works

• Frame and classify. Identify the original, useful decision, simpler alternatives and whether this is a model, shadow or twin.
• Define evidence contracts. Specify identity, schema, clocks, provenance, freshness, uncertainty and replayable model/scenario versions.
• Test before effects. Compare read-only/shadow behavior with independent reality and fail to hold on stale or contradictory evidence.
• Govern the lifecycle. Separate approval, execution, rollback and stop, with scoped credentials, owners and retirement checks.

Full instructions

• Open Claude, ChatGPT or Gemini. Paste in the skill text, or upload the skill folder if your AI accepts files.
• Describe, in this order, the system, its data sources, the decision the twin should help with, the owner, the risk level and the actions it may take. If you do not know one, say so.
• Ask for the classification first, then the plan and the release checklist.
• Read the open gaps and fill them in before you trust the plan. Then ask the named owner to approve, hold or block the release, and act on that decision: approve means proceed, hold means fill the gaps, block means stop.
• No code or paid tool is needed. The plan is a document, not a running twin.

Words used in this task

• Model: A stand-in for a real system that does not update from it.
• Digital shadow: A copy that is fed by the real system but cannot send anything back.
• Digital twin: A copy kept in step with the real system that can help decide or trigger actions, under set rules.
• Digital thread: The linked record that ties data, models and decisions together over time.
• VVUQ: Checks that the model is built right, matches reality, and shows how sure it is.

Full instruction file

Digital Twin

Use this skill to design and operate a trustworthy digital twin, especially a twin of a software product, engineering process, delivery system, infrastructure estate, or agentic factory.

Before you start

Use the material below to do this job. If facts or tool access are missing, explain what you can finish and what must wait.

What you need (use your own material or the linked skill):

• Your question (yours): the decision or problem, and what a good outcome looks like.

Makes it better (optional):

• Source text (yours): a long source in full: an article, a video or talk transcript, or notes from a book, with its title, author and link.

How to start:

1. Look first. Use what they've already attached, pasted or put in their about-me note. In a coding assistant with file access, look in the project too for these by name.
2. Check what's missing. Look for your question, source text. They can attach files, paste the details, or share a file path your assistant can read. Ask only for missing facts that change the result. If they ask you to start with what they have, complete the parts you can and mark the gaps.
3. Do the job. Follow the skill below with the facts and tools you have. Mark each gap with MISSING: where it belongs and never invent a fact to fill it.

Core position

A digital twin is not automatically a dashboard, 3D view, dependency graph, emulator, test fixture, or LLM wrapper. Establish the represented original, synchronization contract, models/services, intended decision, uncertainty, and action boundary before using the term.

For this skill, use these working distinctions:

• Digital model: representation without a required live synchronization loop.
• Digital shadow: observable flow from original to representation without a governed return path.
• Digital twin: a versioned representation of a named entity or process, synchronized at an explicit frequency and fidelity, with models/services that support a declared decision or action and a governed feedback relationship to the original. If it is synchronized and useful but has no governed return path, call it a read-only twin or digital shadow rather than implying closed-loop authority.
• Digital thread: provenance and lifecycle linkage across systems, artifacts, decisions, and outcomes. A thread is necessary for many twins but is not itself a twin.
• Twin universe: a federation or composition of purpose-bounded twins with explicit identity, time, provenance, semantics, ownership, and authority boundaries.

These are operational definitions, not claims of universal standards consensus. NIST IR 8356 explicitly notes that no single definition is agreed, while allowing abstract entities and processes as twin subjects.

Choose an operating mode

Do not start in a higher-authority mode merely because the model is confident or the user calls the system a dark factory.

Choose an entry point

Then follow the numbered workflow and load the matching reference in its navigation table.

Operating workflow

Load the matching reference before acting:

1. Frame the use case. Name the original, decision, desired outcome, authoritative sources, update/fidelity requirement, operating domain, owner, risk tier, forbidden actions, and simpler alternatives.
2. Classify the representation. Decide whether the artifact is a model, shadow, twin, emulator, simulator, or composite universe. Do not upgrade the label without evidence of synchronization, model credibility, and a governed feedback path.
3. Design the contracts. Define stable identities, event envelopes, timestamps, schema versions, provenance, freshness, uncertainty, validity intervals, permissions, and action semantics.
4. Build evidence first. Capture immutable source events and artifacts, then create replayable temporal projections. Preserve raw evidence separately from claims and disposable views.
5. Add models and scenarios. Register model purpose, assumptions, domain, parameters, version, calibration, uncertainty, tests, and prohibited use. Pin scenario inputs, model digests, seeds, clocks, fixtures, and outputs.
6. Run read-only and shadow modes. Compare reconstructed state and recommendations with independently captured reality before allowing side effects. Treat missing, stale, contradictory, or unauthenticated evidence as unknown or hold, not pass.
7. Govern the action path. Separate observe, simulate, recommend, approve, execute, rollback, and stop. Put authorization in deterministic policy enforcement, not model prose. Require idempotency, scoped credentials, preconditions, expiry, rollback, and reconciliation.
8. Evaluate independently. Keep represented-system health, synchronization/data health, model credibility, platform health, and agent/action quality as separate planes. The twin must not grade itself.
9. Operate and evolve. Monitor freshness, loss, ordering, schema/topology drift, residuals, calibration, uncertainty, scenario gaps, platform SLOs, agent trajectories, authority, and outcome value. Revalidate after material changes.
10. Retire deliberately. Revoke action authority, drain and migrate consumers, preserve required lineage, remove secrets and sensitive data under policy, mark endpoints retired, detect orphan calls, and verify no live workflow still depends on the twin.

Reference loading

The Operating workflow navigation table is canonical: load the matching reference there before acting. The same references remain directly addressable for deep dives: architecture, implementation, evaluation, governance, lifecycle, and source-index.

Agentic software factory boundary

A software factory analogue maps repositories, revisions, requirements, builds, artifacts, dependencies, services, environments, deployments, incidents, humans, agents, policies, and runtime observations to twin entities and events. Executable repository environments, service emulators, traffic mirrors, IaC sandboxes, and learned world models are useful components, but each has a simulator-reality boundary. Real execution remains the release oracle for consequential code and infrastructure changes.

A dark factory is an authority state, not a twin type. More automation requires stronger independent validation, provenance, staged rollout, rollback, and accountable escalation. Do not equate high model confidence with authority.

Choose the artifact

Use the smallest artifact that answers the decision; do not create a twin universe when a bounded model, shadow, emulator, or ordinary observability record is sufficient.

Minimum release packet

Before granting any authority beyond read-only, preserve the following evidence packet. It is the release gate, not a suggestion:

• intended-use, risk, and authority contract;
• source and identity map;
• event/schema/provenance contract;
• assumptions, validity limits, and uncertainty budget;
• verification and independent validation results;
• synchronization/data-quality benchmark;
• scenario and replay manifest;
• agent task and trajectory evaluation;
• security, privacy, and supply-chain review;
• shadow/canary and rollback evidence;
• health SLOs and alert ownership;
• signed approve, conditional approve, hold, or block decision.

When not to use

Do not load this skill for ordinary dashboard construction, static asset inventory, generic dependency mapping, generic AI governance, generic DevOps design, or routine operation of a named tool. Route those tasks to the relevant observability, AI-governance, platform, data, or tool-specific skill. In this repository, use agent-evals-and-observability for generic agent eval/telemetry, ai-governance for organization-wide AI governance, data-architect for general data-platform design, and the relevant tool skill for operating a named platform. Load this skill when the representation itself, its synchronization, simulation/prediction, composition, authority, or lifecycle is the problem.

Exit criteria

Stop when the requested twin design or decision artifact exists, claims are separated into evidence and inference, the relevant evaluation/governance gates are explicit, and unresolved gaps have owners or bounded escalation. Never claim a twin is trustworthy, autonomous, production-ready, or safe solely because its schema, graph, dashboard, or component tests pass.

After this

When you've handed it back, suggest what they could do next in one short question. It's a suggestion: if they're done, that's fine. Offer these, and say they can type their own:

• The MATLAB signal filter designer. It takes the audit report. If they pick one and that skill is added here, carry straight on with what you just made. If not, give the link and say what to bring.