Tie Breaker

tie-breaker

Two agents or vendors disagree on how to clear a blocker. This skill does not pick a winner. It restates the disagreement in plain language, synthesizes two middle-ground options, and publishes a self-contained HTML page so you make the call.

Inputs

• The blocker or issue statement.
• Model A: name, recommendation, reasoning (paste text or a file path).
• Model B: name, recommendation, reasoning.
• Model B can be any agent or vendor (Hermes is the worked example in assets/example.json). If a side's position is missing, get it:
  • Another CLI agent (example, Hermes): hermes chat -Q -q "<prompt>" --reasoning medium (one-shot; prints a session_id then the answer).
  • Claude: reason inline, or dispatch an Agent for a fresh independent take.
  • A third opinion from another vendor is optional, never required.

Procedure

1. Restate the blocker in layman's terms. No jargon. Assume a smart non-engineer reader.
2. Present each side's recommendation in plain language: what it does, why that model thinks so, what it costs or risks, what could go wrong.
3. Synthesize exactly two middle-ground options built from the two recommendations (e.g. sequence A then B with a checkpoint, or take A's mechanism with B's safeguard). Each option needs: a one-line summary, ordered steps, tradeoffs, who does what, and how to reverse it.
4. Build a comparison table: effort, risk, reversibility, speed, and which model would object and why.
5. Add a "what I would pick and why" line from Claude, clearly labeled as opinion, not a decision.
6. Render and publish:
   • Fill a JSON file matching assets/example.json's shape.
   • Run node scripts/render-tiebreaker.mjs <input.json> <output.html>, writing the HTML to $OUTPUT_DIR/tie-breaker/<YYYY-MM-DD>-<slug>.html.
   • Publish: run $PUBLISH_CMD <file> (any static-HTML host or internal dashboard works; see Configuration).
   • Report the URL the publish step prints, using a hostname your reviewer can reach from their device, not localhost.
7. Print a short terminal summary (about 5 lines): the two middle-ground options and the published link. Then stop and wait for the human's pick. This skill never executes the chosen option itself.

Writing rules

• No em-dashes anywhere, in the JSON content or the rendered HTML.
• Plain language throughout: this page is for a decision, not a technical audit.
• The JSON is the source of truth for layout; do not hand-write HTML for a real run, always go through render-tiebreaker.mjs so every tie-breaker page looks the same.

JSON contract

See assets/example.json for a complete worked example. Top-level keys:

{
  "issue": "plain-language restatement of the blocker",
  "models": [
    { "name": "...", "recommendation": "...", "reasoning": "...", "risks": "..." }
  ],
  "middle_ground": [
    { "title": "...", "summary": "...", "steps": ["..."], "tradeoffs": "...", "who": "...", "reversal": "..." }
  ],
  "comparison": [
    { "option": "...", "effort": "...", "risk": "...", "reversibility": "...", "speed": "...", "objection": "..." }
  ],
  "claude_pick": "one paragraph, clearly labeled as opinion"
}

models normally has exactly two entries (A and B); a third opinion from Gemini or Codex can be appended as a third entry if one was gathered. middle_ground must have exactly two entries; two is the point of the exercise.

What this skill does not do

• It does not execute the chosen option. It stops after publishing and reporting the link.
• It does not run evals or the skill-creator eval loop.
• It is not a general second-opinion tool; that is a second-opinion skill, which produces one opinion, not an adjudication between two existing ones.

Configuration

This skill expects:

• $OUTPUT_DIR: where rendered tie-breaker pages are written (default: ~/outputs).
• $PUBLISH_CMD: a command that takes one HTML file and prints a shareable URL. Any static-HTML host, internal dashboard, or a plain cp to a served directory works. If unset, the skill reports the local file path and asks you where to publish.
• Optional: a CLI for the second agent (the worked example uses hermes chat). Without one, paste Model B's recommendation directly.