Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
quant-llm-skills
🇺🇸 English | 🇪🇸 Español | 🇨🇳 简体中文
Skills for quant research with LLMs that don't fall for the traps nobody talks about.
Most "AI trading" skill packs teach Claude how to backtest CANSLIM and parse 10-Ks for AAPL. They fail the moment you point them at a real small-cap with sparse XBRL, an active ATM, and four 13D filers reporting the same shares.
🎯 Especially valuable on Haiku and smaller models. Our regression suite shows 8–9 out of 11 evals measurably differentiate Haiku output — caught only with the plugin loaded. Cost-conscious users running cheaper models are silently shipping broken analyses. Reproduce in ~100 seconds: make evals-baseline. See BENCHMARKS.md.This pack distills hard rules from running production quant pipelines into Claude Code skills the LLM applies automatically — no system-prompt hacking, no manual invocation, no extra glue.
What it catches that a baseline LLM misses
- Lookahead bias in any historical query —
period_endis not a
publication date.
- Capacity vs action — an effective S-3 alone is not a dilution event.
- ATM vs registered direct vs ELOC — same 424B5, different trades.
- The placement agent matters — Goldman on a secondary ≠ Wainwright
on an ATM.
- XBRL gaps for FPIs and SPACs — the data is in the filing text,
not the structured feed.
- 13D group double-counting — naive sum of cover pages overstates
insider ownership 2–10x.
- Quantitative dilution scoring — 0–100 with auditable component
weights, not vibes.
See EXAMPLES.md for real prompt-and-response transcripts.
Skills
| Skill | What it does |
|---|---|
| lookahead-safety | Forces filing_date as known-date, never period_end. The #1 quant-backtest bug. |
| sec-filing-types | Disambiguates SEC forms (S-3, 424B, 8-K items, 13D/G, Form 4, 20-F, 6-K, NT 10-K). Knows that a shelf is capacity, not action. |
| atm-detection | Multi-signal inference for active ATMs. Distinguishes ATM from ELOC and registered direct. Catches dilution that 8-K-only scanners miss. |
| bank-tier-classification | 4-tier framework mapping placement agents (bulge bracket → small-cap specialist) to expected deal behavior. |
| xbrl-fallbacks | When SEC XBRL is empty or 404 (FPIs, recent IPOs, SPACs), defines the cover-page hierarchy and extraction rules. |
| dilution-event-scoring | 0–100 framework integrating ATM + agent tier + recency + cash runway + structure + history. Reproducible, auditable, with action thresholds. |
| insider-dedup | Joint-filer / group / family-attribution dedup rules for 13D/G and Form 4 aggregation. Stops the cover-page-sum bug. |
| code-review-for-quant | Domain-specific code-review checklist (lookahead, splits, snapshots, NaN propagation, joint-filer dedup). Ranks bugs by silent-corruption potential, not by severity-of-symptom. |
| transaction-cost-modeling | Realistic friction defaults for small caps. Catches borrow APR fiction (3% on Reg-SHO names instead of 50–500%), locate-failure-as-slippage bugs, and engine-default near-zero friction. |
| survivorship-bias | Catches the universe-built-from-today's-survivors trap. Special focus on small-cap patterns: reverse-split-then-delist phantom returns, ATM-into-delisting, SPAC merger flips. |
The skills compose: ask "score X's dilution risk" and the scoring skill calls the ATM, agent-tier, and lookahead skills automatically.
Install
Once published to GitHub:
/plugin marketplace add jefrnc/quant-llm-skills
/plugin install quant-llm-skills@quant-llm-skillsFor local testing now:
/plugin marketplace add /absolute/path/to/quant-llm-skills
/plugin install quant-llm-skills@quant-llm-skillsOr one-shot via CLI without installing:
claude --plugin-dir /absolute/path/to/quant-llm-skills -p "your prompt"Who this is for
- Quant retail / boutique funds running point-in-time backtests on
small caps.
- Anyone using Claude / Cursor for SEC EDGAR research who has been
burned by lookahead bias or XBRL gaps.
- Trading-system devs who want the LLM to automatically apply
dilution-detection rules without re-explaining them every session.
Who this is NOT for
- Index / ETF / large-cap-only research where structured data is dense
and dilution is rare.
- Pure-fundamental long-term investors uninterested in placement
mechanics.
- Anyone expecting the LLM to predict prices — these skills make
the LLM more rigorous about the data, not clairvoyant about the future.
Tradeoff
These skills bias toward flagging more risk than baseline. False positives on dilution risk are cheap; false negatives are expensive (unexpected printing into your long). Adjust the scoring thresholds in dilution-event-scoring if you want a less conservative profile.
Validate
claude plugin validate /path/to/quant-llm-skillsAll skills pass claude plugin validate and load via --plugin-dir.
Cursor support
Equivalent rules for Cursor live under .cursor/rules/. They are auto-generated from the canonical SKILL.md files via scripts/sync-cursor.py; see CURSOR.md for details.
More
EXAMPLES.md— real prompt-and-response transcriptsBENCHMARKS.md— empirical Haiku / Sonnet / Opus comparisonsevals/README.md— regression test suite (one per skill)LAUNCH.md— launch playbook (publish, post, iterate)CHANGELOG.md— version historyCURSOR.md— Cursor usage and sync workflow
Contribute
Issue templates for skill misses and skill requests live under .github/ISSUE_TEMPLATE/.
Common tasks:
make validate # static checks
make sync # regenerate .cursor/rules/ from skills/
make smoke # live dispatch test (requires claude CLI + auth)
make evals # regression eval suite (Haiku, ~$0.05/run)
make evals-baseline # evals + baseline (no plugin) comparisonLicense
MIT