Week 1 Post-Mortem: What My Claude Investment Skills Got Right, Wrong, and Missed Entirely
I have been running six Claude Code skills on my real portfolio for the past week. Here is an honest accounting of what happened — the correct calls, the wrong ones, and the signals I should have acted on but didn't.
No survivorship bias. All four cases, in order.
Signal 1: TEM $47.13 Stop-Loss — The Right Call for the Wrong Reason
What the system flagged: Price Monitor hit the stop-loss threshold at $47.13. Alert fired to Telegram at 9:34am. I exited the position.
What happened after: TEM continued lower to $43.80 before stabilizing. The exit was correct on price.
The honest part: I was holding TEM partly because the P/C ratio had read bullish (0.50) a few days before the exit. That reading had not changed materially. The stop-loss fired on price action, not on options signal — and that was right. Price was telling me something the P/C ratio wasn't.
Verdict: Correct exit, but I was relying on the wrong signal. The stop-loss automation worked exactly as designed. The options reading was noise in this case. If I had been watching only the P/C and not the price monitor, I would have held through a 7% additional loss.
What I changed: Stop-loss triggers on price. Options flow informs position sizing decisions, not exit triggers. These are different tools.
Signal 2: XLI P/C 5.32 — Correct Read, No Trade
What the system flagged: XLI put/call ratio of 5.32, classified [OUTLIER — INSTITUTIONAL HEDGE SIGNAL]. The scanner noted this as 4x+ the baseline, consistent with institutional downside protection buying on industrials.
What happened after: XLI dropped 2.3% over the following four sessions. Not a dramatic move, but directionally the signal was right — there was genuine downside pressure on industrials that week.
The honest part: I didn't trade on it. My rule is one data point doesn't make a thesis. I flagged it, noted it in the journal, set a monitoring trigger. The sector moved the way the signal suggested and I was on the sideline.
Verdict: Correct signal, no position, no gain. I'm at peace with this one. A single anomalous reading shouldn't move capital. The system caught something real; the process prevented me from acting on a single data point. That's the right behavior.
One thing I'd do differently: I should have used this to reduce sizing on any existing industrial-adjacent exposure. I had one position adjacent to the sector (small, unrelated thesis) and didn't reduce it. That's a gap in how I connect signals to existing positions.
Signal 3: RXRX P/C 0.38 — Wrong Read, Real Loss
What the system flagged: RXRX raw P/C 0.38, classified [EXTREME BULLISH]. I entered a small position.
What happened after: RXRX fell 3% the next session and continued lower. I honored the stop and exited at -9.9% on the position.
The honest part: I wrote about this separately, but the short version: I did not run the lottery call filter before entering. Adjusted P/C was 2.14, not 0.38. The raw signal was noise. 91.4% of the call volume that made it look bullish was retail OTM lottery tickets.
This was a workflow failure, not a signal failure. The tool had the filter. I skipped it because I was in a hurry.
Verdict: Wrong. My fault, not the system's. The loss was small and within the stop I had set. But the root cause was me bypassing a step I had built specifically to prevent this kind of error.
What I changed: Lottery call filter is now mandatory. The scanner will not output a P/C classification without running the filter first. I removed the option to skip it.
Signal 4: IREN $56.47 — Missed the Bounce
What the system flagged: IREN P/C 0.83 at the time of the earnings miss. The scanner classified it [NEUTRAL]. No alert, no action.
What happened after: IREN reported earnings that missed estimates. Stock dropped 8% after hours. Then reversed — hard. Over the next three sessions it recovered to $56.47, above the pre-earnings price. Anyone who bought the dip made 14%+ in three days.
The honest part: My scanner completely missed this. P/C 0.83 is neutral — nothing actionable. But if I had been running an earnings surprise scanner alongside the options flow scanner, I might have flagged the post-miss setup. Biotech and energy names with high short interest often see violent reversals after earnings dips. IREN has both characteristics.
Verdict: Missed opportunity. Legitimate gap in the system. This isn't a signal failure — it's a feature I don't have. Post-earnings reversal detection requires a different kind of analysis: short interest, float, historical earnings reaction, days-to-cover. None of that is in the current skill set.
What I'm doing about it: Adding an earnings calendar layer to the morning briefing. Not full post-earnings reversal detection yet, but at minimum flagging names that reported within the last 3 sessions so I can evaluate manually.
Scoreboard: Week 1
Signal Result My Call System Fault?
─────────────────────────────────────────────────────────
TEM $47.13 exit Correct Correct No
XLI 5.32 hedge Correct No trade No (correct behavior)
RXRX 0.38 long Wrong Wrong No (I skipped the filter)
IREN bounce Missed No signal Yes (feature gap)
One clean win. One correct non-trade. One self-inflicted loss. One genuine gap.
That's probably about right for week one of any system.
What This Week Taught Me
The stop-loss automation is the most reliable part. It does exactly what it's supposed to do without hesitation. I have hesitated on manual stops before; the automation does not.
Raw P/C ratios without the lottery filter are actively misleading. This is not a minor adjustment — it's the difference between a bullish read and a bearish one on a retail-heavy name. The filter should never be optional.
Signals don't connect to existing positions automatically. XLI flagging sector hedge risk while I held an adjacent position is a workflow gap. The scanner sees the signal; it doesn't see my book. That connection is still manual.
There are whole categories of edge the system doesn't have yet. Post-earnings reversals, short squeeze setups, catalyst calendars. These are additions, not fixes — the current system does what it was built to do, but there's more to build.
Week 2 starts Monday. I'll do the same accounting next week.
The skills used in this post-mortem: github.com/tellmefrankie/ai-investment-skills
Full suite with the stop-loss monitor and options flow analyzer: jaehyunpark.gumroad.com/l/tcyahy
Not financial advice. This is a personal tooling post-mortem, not investment recommendations. All figures are from my own portfolio.
Top comments (0)