Agents

Claude skill

The MCP server gives Claude the tools; the skill gives it the judgment: always price-check before spending, confirm with the user before a paid run, read most_likely_outcome and the distribution instead of over-trusting the single-market verdict, respect the degraded and consensus_market_conflict flags, and cite the public hit rate when presenting confidence.

Install

git clone https://github.com/daveb1392/predictioncouncil-mcp
cp -r predictioncouncil-mcp/skills/predictioncouncil ~/.claude/skills/

That's it — Claude Code picks it up automatically. The skill triggers when you ask Claude to analyze a prediction market, get AI consensus on an event's probability, or compare market odds to model estimates.

What it enforces

## Workflow (always in this order)
1. Find the event      — search_events (free)
2. Check before spend  — check_report: dynamic quote + cache status. NEVER skip.
3. Confirm with user   — if the run costs credits
4. Run                 — run_council_report (30-120s, polls for you)
5. Read the verdict    — most_likely_outcome, market_probabilities,
                         top_bets (edge vs pick), provider_failures/degraded

## Interpreting results honestly
- Cite the public hit rate from /api/stats/track-record
- SKIP on the queried market is common; the distribution still informs
- Analysis, not financial advice — say so

Use both together

Skill + MCP server is the intended pairing: install the server for capability, the skill for discipline. Agents without the skill tend to skip the price check and misread SKIP verdicts — the two mistakes that waste the most credits.