Show and tell: options flow skill that caught XLI P/C = 5.32 live (free, SKILL.md)

[tellmefrankie]

Sharing a Claude Code skill that's been in production for 6 months on a real portfolio.

What it caught:
XLI (Industrial ETF) put/call ratio at 5.32 while SPY was at 0.44. That's 5.96x XLI's own 30-session baseline — institutional-scale downside hedging on industrials while tech longs held. The divergence was the signal.

Why raw P/C ratios fail:
CEG showed ~1.0 (neutral) but 98.4% of its options volume was $0.01–$0.09 lottery calls. Including those in the P/C calculation poisons the signal. One filter line in the skill prompt changes the entire read.

Skill architecture (SKILL.md format):

1. Fetch options chain per ticker
2. Strip contracts ≤ $0.10 (lottery filter)  
3. Compute adjusted P/C ratio
4. Compare vs 30-session rolling baseline
5. Output: ticker, raw PCR, adj PCR, deviation multiplier, signal label

Two skills are free and open source:

• Options Flow Analyzer
• News Sentiment Engine (RSS → Claude → Telegram)

GitHub: https://github.com/tellmefrankie/ai-investment-skills

Works with Claude Code, Cursor, Codex CLI. Happy to discuss the skill architecture.

[affaan-m]

Thanks for sharing this. The useful ECC-facing part is the concrete SKILL.md packaging pattern and the fact that it has a real workflow behind it.

For readers evaluating or adapting it, I would keep a few boundaries clear:

• this should be treated as a community example, not financial advice or an ECC-endorsed trading signal;
• document required data providers, API keys, rate limits, and where credentials are stored;
• include backtest or replay evidence if you want people to trust the signal quality;
• keep any Telegram/provider secrets out of the repo and out of skill examples;
• make the skill output explicit about uncertainty, source timestamps, and assumptions.

If this evolves into a generalized skill-quality example, the most useful follow-up would be a small eval fixture showing how the skill handles noisy options chains, stale data, and missing provider responses.

[affaan-m]

Love this — the live XLI P/C catch is a nice concrete proof point — turning the SKILL.md into a tradeable signal pipeline is exactly the kind of cross-domain use we want featured.

A few things from our side:

1. Sponsor tier refresh just went live — if ECC saves you time enough that you'd like your name in the wall of sponsors, the Builder tier is $25/mo (was $10) and gets you listed in SPONSORS.md plus a private monthly progress note. No pressure on the pitch — just making it easy to find.

2. Happy to feature this in a Show & Tell roundup on our next release if you're up for it — I'd link your repo/gist/post in the release notes and tweet from @affaanmustafa. DM works or reply here.

3. Keep building.

— Affaan

[tellmefrankie]

Thanks Affaan! Would love to be featured in the Show & Tell roundup.

Here's what we've built since posting:

• Lottery-ticket filter — strips $0.01 calls from P/C ratios. RXRX went from 0.38 (bullish) to 2.14 (bearish) after filtering 84% lottery volume
• 9-agent investment team — Scanner, Critic, Analyst, Strategist, Cross-Validator debating positions via threaded Discussions
• Anti-Narrative Harness — forces agents to cross-validate data, no panic selling, numbers over stories

Live catch: XLI P/C hit 5.32 for 5 straight weeks while SPY sat at 0.44. Scanner flagged the industrial hedging divergence.

Repo: https://github.com/tellmefrankie/ai-investment-skills
SKILL.md format for each skill.

Happy to be in the roundup — DM works or reply here. Keep building.

[affaan-m]

Appreciate the update. I’ll treat this as a community Show & Tell candidate for the ECC 2.0 release roundup.

For the roundup I’ll frame it as a skill-packaging / multi-agent workflow example, not as an endorsed trading signal. The interesting ECC angle is:

• concrete SKILL.md packaging;
• explicit filters and assumptions in the tool output;
• agent roles debating and cross-validating a decision;
• a real repo people can inspect and adapt.

The anti-narrative harness piece is especially relevant to the broader ECC direction: agents should preserve source timestamps, uncertainty, and contradictory evidence instead of turning one live signal into a confident story. If you add a small replay/eval fixture for one noisy options-chain case, that would make it much easier to feature cleanly.

[tellmefrankie]

Replay/eval fixture ready.

Two test cases:

1. rxrx-noisy-options.json — 84% lottery calls, raw P/C 0.38 (bullish) → adjusted 2.38 (bearish). Signal completely inverts.
2. cel-clean-options.json — 4.8% lottery, raw ≈ adjusted. No inversion. Control case.

Test suite (test/lottery-filter.test.ts): 12 assertions, all passing. Covers signal inversion detection, lottery percentage thresholds, and contrast between noisy vs clean chains.

Also updated the scanner to output raw + adjusted + lottery composition together (per @MrTalecky's feedback on elizaOS — both figures flow to the Critic now, not just the cleaned number).

All in the repo: https://github.com/tellmefrankie/ai-investment-skills

[tellmefrankie]

Hi @affaan-m — circling back with the piece you asked for. The replay/eval fixtures are live: rxrx-noisy-options.json (84% lottery calls, raw P/C 0.38 → 2.38 after filtering — the signal inverts) and cel-clean-options.json as the control, with 23 assertions on the signal/reasoning separation — exactly the anti-narrative-harness angle you flagged.

Congrats on shipping ECC 2.0 — completely understand if the roundup was already locked by then. If there's a future Show & Tell, or a fit for the harness / skill-packaging examples, I'd be glad to have it considered.

Since we last talked it's had steady organic pickup — regular clones, and a paid user who found the bundle straight from the repo — so the "real repo people can inspect and adapt" angle is holding up. No pressure either way, and thanks again for the thoughtful framing earlier.