Synthesis

• Read synthesis/graph/edges.yaml
• Find all edges of type amplifies: Which concepts create disproportionate impact together?
• Form chains: A amplifies B amplifies C → three-stage strategy
• Per chain: Describe total effect, summarize prerequisites

2. Cross-Impact Analysis

• Select the top-20 concepts by edge count from synthesis/graph/nodes.yaml
• For each pair: How does concept A affect the impact of concept B? (-3 to +3)
  • +3 = A strongly amplifies B
  • 0 = no effect
  • -3 = A strongly inhibits B
• Calculate per concept:
  • Active sum = sum of row values (lever: high = strong influencer)
  • Passive sum = sum of column values (indicator: high = strongly influenced)
  • Criticality = active × passive (high = both influences and is influenced)
• Save result as synthesis/graph/views/cross-impact.md

3. Calculate combinations

For each promising 2-way and 3-way combination:

1. Identify promising pairs from synergy chains and high-criticality concepts
2. Per combination, create a worked example with concrete numbers
3. If the project has domain agents (e.g., @steuerberater): have all relevant agents evaluate in parallel
4. Document conflicts between perspectives
5. Assess: prerequisites, mechanism, risks, related decision trees
6. Save as synthesis/combinations/combi-XX-name.md

Use the template from ${CLAUDE_PLUGIN_ROOT}/templates/kombination.template.md.

4. Derive Decision Trees

Per main decision (e.g., "Which strategy to pursue?"):

1. Identify the central question
2. Build if-then structure with branches
3. Reference combinations and sources at each leaf
4. Document assumptions and limitations
5. Save as synthesis/decision-trees/dt-XX-name.md

Use the template from ${CLAUDE_PLUGIN_ROOT}/templates/decision-tree.template.md.

5. Construct argument chains

When synergy chains or combination strategies reveal multi-step reasoning (A → B → C → conclusion), construct formal argument chains:

1. Build the chain — for each link in the reasoning:
   • State the warrant (why does A lead to B?)
   • Assess the link's qualifier (how confident are we in this specific step?)
   • Identify the evidence tier for this step (1 = meta-analysis/RCT, 2 = observational, 3 = mechanistic, 4 = anecdotal)
2. Compute composite qualifier — chain confidence = min(link qualifiers). A chain is only as strong as its weakest step
3. Identify weakest link — explicitly flag the bottleneck; this is where research effort should focus
4. Classify provenance:
   • literature — chain found as a complete argument in published sources
   • synthesis — constructed by connecting findings from multiple sources/disciplines
   • hypothesis — proposed without direct evidence, based on mechanistic plausibility
5. Check for alternative chains — if two chains explain the same observation, document both and compare composite qualifiers
6. Track disciplines — record which domains each chain crosses; chains crossing 3+ disciplines represent the most novel insights

Save argument chains to synthesis/graph/chains.yaml. Use the template from ${CLAUDE_PLUGIN_ROOT}/templates/argument-chain.template.yaml.

6. Update research questions

• What remains open after synthesis?
• What new questions did the combination analysis raise?
• Which weakest links in argument chains need targeted research?
• Update synthesis/gaps/research-questions.md

Outputs

• synthesis/combinations/combi-*.md
• synthesis/decision-trees/dt-*.md
• synthesis/graph/chains.yaml (argument chains)
• synthesis/graph/views/cross-impact.md
• synthesis/gaps/research-questions.md (updated)
• synthesis/runs/DATE-synthesis/run.yaml + new-combinations.md + insights.md

Specific Combination: /synthesis A+B

When invoked with specific concepts (e.g., /synthesis location-strategy+menu-pricing):

1. Verify both concepts exist as nodes in the graph
2. Check existing edges between them
3. Calculate the specific combination using the template
4. If domain agents exist, have them evaluate
5. Save as synthesis/combinations/combi-XX-name.md

Delegation to @synthesizer

For analytical tasks (graph traversal, pattern recognition, cross-impact calculation), delegate to the @synthesizer agent. The agent handles methodology; this skill handles orchestration and file output.

Important

• All outputs in English
• YAML must be valid (observe required fields, see templates)
• Each run creates an entry under synthesis/runs/
• Worked examples with concrete numbers, label assumptions clearly
• Decision Trees with source references
• Distinguish between facts (from graph) and interpretations (from analysis)