WIN Committee

Overview

Execute the end-to-end committee pipeline for hackathon and MVP messaging validation.

Use this orchestration skill when the user wants full-loop output (profiles.json, committee_matrix.json, and summary.md).

For single-stage tasks, prefer focused skills:

• •win-profile
• •win-evaluate
• •win-summary

Workflow

1. •Validate inputs against the contracts in references/data-contracts.md.
2. •If required inputs are missing, please interview the user to acquire the necessary inputs before running commands.
3. •Run profiling from target public data.
4. •Run committee evaluation across experts and content drafts.
5. •Synthesize output into a balanced summary with consensus signals.

Commands

Run full pipeline:

scripts/run_committee.sh

Use the Laura evidence-rich example as input:

cd /Users/tomago/andrew-tomago/public/weighted-intelligence-nodes
cp examples/targets.laura-modiano.example.json data/input/targets.local.json

Run staged pipeline:

python3 ../../../scripts/mvp_committee.py profile \
  --targets ../../../data/input/targets.local.json \
  --out ../../../data/output/profiles.json

python3 ../../../scripts/mvp_committee.py evaluate \
  --profiles ../../../data/output/profiles.json \
  --content ../../../data/input/content.local.json \
  --committee ../../../config/committee.json \
  --out ../../../data/output/committee_matrix.json

python3 ../../../scripts/mvp_committee.py synthesize \
  --matrix ../../../data/output/committee_matrix.json \
  --out ../../../data/output/summary.md

Tuning Rules

• •Add or remove committee experts in config/committee.json.
• •Adjust score emphasis via rubric_weights.
• •Define judging_criteria labels/descriptions in judge language to improve persona realism.
• •Keep weights normalized enough to avoid one expert dominating output.
• •Add domain-specific focus keywords per expert to reflect committee specialization.
• •Check committee_matrix.json rationale for empathy and specificity before sharing summary.

Guardrails

• •Use only public or user-provided data.
• •Never infer private or sensitive attributes.
• •If required inputs are missing, please interview the user to acquire the necessary inputs.
• •Keep analysis focused on content quality and audience-fit signals.
• •Stick to observable public work and stated preferences only.
• •Keep user-specific input in data/input/*.local.json (gitignored).
• •Keep all file paths repository-relative or skill-relative; avoid host-specific absolute paths.

References

• •references/workflow.md
• •references/data-contracts.md
• •scripts/run_committee.sh