prediction-market-live-ops — Claude Code skill
prediction-market-live-ops — a skill for automated trading on prediction markets (Kalshi). Combines backtest methodology, live-execution safety rails, and hypothesis-falsification protocols derived from a real live-money campaign.
MIT 194 installs ★ 582 updated 2026-09-03
npx skills add agiprolabs/claude-trading-skills --skill prediction-market-live-ops -a claude-codeWhat prediction-market-live-ops does
This skill packages operational knowledge from a live Kalshi trading run (August–September 2026: crypto/commodity maker ladders, regime tilts, tennis dust). Every strategy decision requires evidence: backtest → paper trading → live micro-probe → scale. Each stage has pre-declared kill conditions with properly calibrated probability math, written down before data arrives. If a threshold proves miscalibrated, correct it openly and re-declare.
The skill covers backtest hygiene (print-replay as an upper bound, correct time-series order, fill-conditioned validation, PnL concentration checks, early-settlement bias), live execution rules (smoke-test every order path, distinguish cancelled from filled orders, consistent shard routing, empirically measured rate limits, post_only for makers, halt-file and drawdown breaker, queue priority, geofencing), and scaling doctrine (house-money ratchet, capacity-curve measurement, fill-quality gates).
Built on reference implementation at https://github.com/agiprolabs/kalshi-stack and complements prediction-market-strategy (edge selection and sizing). This skill handles the operational side: validating, executing, and halting strategies on a live venue such that losses stay within the cost of the lesson.
What it helps with
- Design a prediction-market backtest that avoids phantom profits from adverse selection
- Deploy a trading bot on Kalshi with position-crash protection against rate limits and race conditions
- Distinguish regime drift from a real edge when a pattern holds in one time period but vanishes in another
- Scale a strategy in steps while maintaining the correlation between size and realized PnL
- Maintain append-only fill ledgers and falsification records for post-hoc audit and result trust
Example prompts for the agent
- Design a backtest for a maker strategy on tennis dust markets, accounting for print-replay overstatement
- Set up a smoke test for live Kalshi orders and confirm that cancellations properly distinguish filled from cancelled state
- Check whether a pattern across morning hours is a regime bet or true edge if it disappears in the afternoon
How to install prediction-market-live-ops
1Skills CLI (Vercel)
npx skills add agiprolabs/claude-trading-skills --skill prediction-market-live-ops -a claude-code2Manually: put the files in place
~/.claude/skills/prediction-market-live-ops/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/prediction-market-live-ops/
1Skills CLI (Vercel)
npx skills add agiprolabs/claude-trading-skills --skill prediction-market-live-ops -a codex2Manually: put the files in place
.agents/skills/prediction-market-live-ops/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add agiprolabs/claude-trading-skills --skill prediction-market-live-ops -a cursor2Manually: put the files in place
.cursor/skills/prediction-market-live-ops/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install agiprolabs/claude-trading-skills skills/prediction-market-live-ops2Manually: put the files in place
.github/skills/prediction-market-live-ops/SKILL.md1Skills CLI (Vercel)
npx skills add agiprolabs/claude-trading-skills --skill prediction-market-live-ops -a gemini-cli2Platform CLI
gemini skills install https://github.com/agiprolabs/claude-trading-skills --path skills/prediction-market-live-opsManually: put the files in place
.devin/skills/prediction-market-live-ops/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add agiprolabs/claude-trading-skills --skill prediction-market-live-ops -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about prediction-market-live-ops
How do I install prediction-market-live-ops in Claude Code?
Run npx skills add agiprolabs/claude-trading-skills --skill prediction-market-live-ops -a claude-code, then open ~/.claude/skills/prediction-market-live-ops/SKILL.md.
What does prediction-market-live-ops do?
It provides methodology and operational rules for building, backtesting, and running automated strategies on prediction markets like Kalshi. The skill covers honest historical validation, live execution with safety systems, and staged scaling.
How is this different from prediction-market-strategy?
prediction-market-strategy helps you choose an edge and size a position. prediction-market-live-ops covers the operations: validating a strategy honestly, launching it live without blowups, halting if outcomes diverge from expectations.
Source text
SKILL.md · MIT
--- name: prediction-market-live-ops description: Use when building, backtesting, operating, or scaling automated trading on prediction markets (Kalshi or similar) — evidence-gated methodology, live-execution safety rails, and falsification protocols distilled from a real live-money campaign --- # Prediction-Market Trading Doctrine Hard-won operational knowledge from a live Kalshi campaign (Aug–Sep 2026: crypto/commodity maker ladders, regime tilts, tennis dust vacuum — all falsified honestly, infrastructure open-sourced). Reference implementation: https://github.com/agiprolabs/kalshi-stack. Complements `prediction-market-strategy` (edge selection/sizing) — this skill covers the OPERATIONAL side: validating, executing, and killing strategies against a live venue without losing more than the lesson costs. ## The prime directive Every strategy decision is evidence-gated: backtest → paper → live micro-probe → ratchet. Each stage has a PRE-DECLARED kill condition with correctly calibrated probability math, written down BEFORE the data arrives. When a threshold turns out miscalibrated, correct it openly and re-declare — never silently move a goalpost in either direction. ## Backtest hygiene (each rule was violated once, expensively) - **Print-replay EV of passive-fill strategies is an UPPER BOUND, never an estimate.** The prints a resting order actually captures are an adversely-selected sample of the tape (sweeps deep enough to reach your queue skew toward true collapses). A tennis dust book that measured +125%/$ on prints went 0-for-78 live (P<2%). Only fill-conditioned live outcomes validate a maker strategy. - **Time-series joins use the last bar ending ≤ the decision instant.** A candle containing the instant is look-ahead and will fabricate covariates that vanish out-of-sample. - **Taker backtests need print-confirmed executability**, not quoted-ask edges — stale quotes create phantom fills. - **Any bucketed-PnL review includes a top-N-share column.** Lottery-payoff books concentrate EV; a "pattern" that is two elephant rounds is noise. Same rule when a holdout "wins": check concentration before believing it. - **Early settlements are selection-biased toward losers** (collapses end fast; comebacks are long matches/rounds). Never read a verdict off the fast-settling prefix. - **Static tilts that flip sign between tape halves are regime bets, not edges.** A conditioning scheme must beat baseline on BOTH halves and on each individual day; pre-register the rules, no per-day fitting. - Mirror/complement markets are the SAME bet — deduplicate before computing independence-based probabilities. ## Live execution rules (the incident ledger) - **Smoke-test every live order path with a real 1-lot round trip** (place, verify resting, cancel, verify gone) before any strategy trades through it. Paper cannot exercise venue routing parameters. Cost: $0.01. - **Cancels must be truthful**: distinguish cancelled / already-gone (filled!) / failed. A 404 is not a cancel — it means wrong routing shard or a fill-race. Swallowed cancel failures + re-anchor loops = runaway position stacking (cost us $84 in one evening). - **Route the exchange shard (`exchange_index`) consistently on EVERY call** — placement, reads, cancels, collateral. Collateral must be pre-positioned on the shard before orders. - **Measure rate limits empirically with dust orders** (burst ladder + sustained test). Advertised budgets can be 10× off effective per-order cost. Pace bursts to the measured rate with retry-requeue on 429. - **post_only for any maker strategy** — an "aggressive maker" price on a market trading inside your band silently becomes a taker fill of a different (unvalidated) trade. Tag and cohort maker vs taker fills separately in all realized-EV accounting. - **Rails, always**: halt-file checked before every order batch (cancel-all on appearance), independent scheduler-driven drawdown breaker with deposit-jump detection, per-order and per-market size caps with anomaly ledger events, equity floor stand-down. The breaker baseline is set at session start per the operator's stated loss tolerance. - **Queue priority is the product** in penny/maker books. Synchronized-open venues: NTP + fire at boundary+0 with a retry probe (measure the venue's open-transition latency). Event-driven venues: push-based discovery (lifecycle WebSocket channels) beats polling; a fresh market's book is empty, so first-to-rest owns the level for its lifetime. - **Restart daemons at activity boundaries**; mid-round restarts orphan in-memory state (untracked orders, missed settlements). - **Geofencing is real**: order-origin IP matters, categories differ (sports/elections vs financials), and enforcement can change overnight. Datacenter location is part of your compliance posture. ## Scaling doctrine - **House-money ratchet**: double size only when realized profit at the current step covers the next step's incremental risk AND round-count and EV-continuity gates pass. Drop back a step on trailing degradation. - **Measure the capacity curve** (EV/$ by per-print cap) before believing any scale projection; then the binding constraint is capture rate × measured pool, and capture rate is only measurable live. - **Fill quality is the untested link at every new size** — the paper/live EV-per-$ ratio at each step gates the next. ## Experiment operations - Cohort every fill (maker/taker, band, source) at write time — post-hoc attribution is what lets a bug contaminate a verdict. - Ship one variable at a time; changing size and placement logic in the same window destroys attribution. - Run A/B variants in paper against a frozen baseline (small-footprint variant, shared rounds, EV/premium-$ as the scale-free judge). - Log everything to append-only ledgers; reconcile realized PnL to the venue's own fills/settlements, never to internal marks. - Post results (wins AND falsifications) to the operator's channel; write a decision record with evidence and reversal conditions for every ship, retire, and kill.
© agiprolabs · MIT · agiprolabs/claude-trading-skills