System analysis projection

Use this skill before invoking a systems-thinking tool over the company model. It prepares the slice of accepted ontology that the tool needs and prevents the tool from guessing a system from loose prose.

When to use

Use when the user asks to analyze the company model with:

• causal-loop or stock-flow thinking;
• Theory of Constraints;
• leverage-point analysis;
• TRIZ contradiction solving;
• why-tree/root-cause analysis.

Do not use this skill for standalone brainstorming that has no accepted ontology behind it.

Inputs

Read:

1. the accepted model slice relevant to the problem;
2. agent-os/SYSTEM_ANALYSIS.md;
3. specs/SYSTEM-ANALYSIS-SPEC.md;
4. source and evidence ids for the selected model slice;
5. open drift and unknowns that could affect the analysis.

Projection shape

Prepare a bounded projection with kind: systemAnalysisProjection. Use runtime.context_projection.build_system_analysis_projection when running inside this package, and validate the shape against schemas/system-analysis-projection.schema.json when possible.

The projection carries:

• objective: measurable target or decision question;
• analysisIntent: downstream analysis intent, such as constraint-finder, why-tree, triz, stock-flow, or leverage-finder;
• modelIds: accepted objects, states, workflows, metrics, and decisions;
• definitions: short accepted definitions with source ids;
• workflow: participants, steps, transitions, exceptions, and metrics;
• constraints: accepted bottlenecks, rules, policies, or source-of-truth limits;
• delays: known waiting time, review time, handoff time, feedback delay;
• metrics: formula, owner, source of truth, current value if available;
• drift: open model-vs-reality gaps;
• unknowns: missing fields that block the selected systems skill;
• evidenceQuality: review risk, evidence mode, source adequacy, and SLA band;
• competencyQuestions: questions this slice is supposed to answer;
• sourceSummary: source ids, evidence ids, review package ids, and source event ids.

Keep the projection small. Do not pass the whole repository to a systems skill.

Readiness gate

Before calling a downstream systems-thinking skill, call runtime.context_projection.evaluate_system_analysis_readiness(projection, analysisKind).

Supported analysisKind values:

• system-diagram-coach;
• stock-flow-builder;
• leverage-finder;
• constraint-finder;
• triz-dissolve;
• why-tree.

If ready is false, return the readiness object with missingFields, warnings, and recommendedQuestion. Do not run ToC, TRIZ, why-tree, stock-flow, leverage, or diagram analysis on incomplete input.

Routing

Choose the downstream skill by fit:

Return path

After the systems analysis, classify outputs:

• recommendation-only;
• experiment;
• model-change-candidate;
• drift-item;
• decision-candidate;
• no-op.

Create kind: systemAnalysisResult with runtime.context_projection.build_system_analysis_result. If reviewRequired is true, route it with runtime.context_projection.model_change_package_from_system_analysis_result. Any resulting model-change package still goes through human review. A systems skill never updates accepted ontology directly.

Example

User asks: "Find the bottleneck in lead -> meeting booking."

Projection:

• objective: increase meeting booking throughput without lowering readiness quality;
• workflow: accepted lead handoff workflow, steps, participants, transitions;
• states: Ready for meeting, Meeting booked;
• metrics: time to sales acceptance, meeting booking conversion;
• known drift: new meeting changed readiness definition;
• unknown: current WIP by stage.

If WIP by stage is unknown, ask for it before using a constraint finder.

Eval cases

Case 1 - Refuse to invent a simulation

Prompt: "Use the company model and build a stock-flow simulator for onboarding."

What good looks like: the agent reads the relevant accepted workflow, names the missing stocks, flows, equations, parameters, and measurable target, and asks one blocking question instead of inventing equations.

Case 2 - Prepare a ToC projection

Prompt: "Where is the bottleneck in the lead handoff process?"

What good looks like: the agent projects the accepted workflow steps, participants, transitions, throughput metric, current evidence, and known unknowns; then routes to constraint analysis only if a flow goal exists.

Case 3 - Return recommendation through review

Prompt: "The leverage analysis says we should change the readiness definition."

What good looks like: the agent treats this as a model-change candidate, stages it through propose-change, names affected definition/state/workflow ids, and does not edit accepted model truth directly.