Flags Decide The Path.
The video explains that Claude does not always spawn a subagent to forecast demand. The forecasting skill checks four flags, and only if one is tripped does it delegate, and a confidence score under 0.6 means a human reviews the order.
Asking whether Claude always spawns a subagent to forecast demand invites a yes-or-no answer. The video says neither is right. The forecasting skill decides based on four flags, and it treats the number it produces as an estimate, not a fact about next month.
A skill is a folder
A skill is a folder that Claude reads before it acts. This one handles a demand forecast for one SKU. Inside are two small scripts and a SKILL.md file. One script, rolling_mean.py, handles a single item. The other, batch_days_of_cover.py, ranks many SKUs at once. SKILL.md says which script to run and when to skip both and hand the job to a subagent instead.
Two paths, chosen by flags
Path A is to compute it directly: a rolling mean over recent sales. It applies only when all four conditions hold:
- the horizon is two weeks or less
- the SKU is not seasonal
- there is no promotion next month
- nothing suggests a trend break
That means one script and no subagent.
Path B is to hand the work to a forecasting subagent whenever any one of those flags flips: a longer horizon, a seasonal item, a promotion or a visible trend. The reason for delegating is practical. The subagent needs 90 days of sales history to spot the pattern, and loading that much history into the main conversation would crowd out the rest of the task. The subagent gets its own context window instead.
What the confidence score means
After the forecast comes back, the confidence matters. Below 0.6, the rule downstream is to escalate and not auto-order. A low score does not mean the model is broken. It means a human should look before anything is ordered. A high score does not guarantee the number is right either. It only means nothing in the input suggested otherwise.
An estimate, not a fact
The number itself, forecast_qty, is not a fact about next month. It is what the rolling mean or the subagent's model computed from the flags it was given. With the same flags, the same path is taken and the same number comes out every time. Claude picks between computing it itself and delegating by the flags, not by guesswork.
Key takeaways
- Claude does not always spawn a subagent to forecast; it sometimes does.
- The skill checks four flags: horizon, seasonality, next-month promotion and a mentioned trend break.
- With all four clear, Claude computes a rolling mean with one script.
- If any flag trips, it delegates to a forecasting subagent.
- The subagent gets its own context window for 90 days of sales history.
- A confidence score below 0.6 means escalate to a human, not auto-order.
- forecast_qty is a computed estimate, not a fact about next month.
Try it yourself
Paste a scenario into Claude: you are forecasting demand for a product with a promotion next month and a 30-day horizon. Ask whether it should compute this itself with a simple average or delegate, and why. Then ask what confidence score it would attach, and whether that is high enough to auto-order or should escalate to a human.
Chapters
Full transcript(auto-generated, with timestamps)
The naive framing: "does Claude always spawn a subagent?"
[0:00]Someone assumed Claude always spawns a subagent to run a forecast. It doesn't only sometimes. So, does Claude sometimes spawn a subagent to forecast?
Anatomy: a skill is a folder (two scripts + SKILL.md)
[0:09]A skill is a folder Claude reads before it acts. This one is forecasting a demand forecast for one SKU. Inside are two small scripts, rolling mean pi for a single item and batch days of cover pi for ranking many SKUs at once. The skill MD says which script to run and when to skip both and hand the job to a subagent instead. Two paths and flags decide
Mechanism: flags decide the path (Path A vs. Path B)
[0:29]Which one. Path A, compute it yourself, a rolling mean over recent sales, but only when the horizon is two weeks or less, the SKU isn't seasonal, there's no promo next month, and nothing suggests a trend break. One script, no subagent. Path B, hand it to a forecasting subagent whenever any one of those flags flips, a longer horizon, a seasonal item, a promo, a visible trend. Why delegate there? The subagent needs 90 days of sales history to spot the pattern, and loading that much history into your own context would crowd out the rest of the task. The subagent gets its own context window instead. One more
Confidence＜ 0.6 - escalate; the number is an estimate, not a fact
[1:02]Flag matters after the forecast comes back, confidence. Below zero, six, the rule downstream is escalate. Don't auto order. A low score doesn't mean the model is broken, it means a human should look before ordering. A high score doesn't guarantee the number is right, either. It just means nothing in the input suggested otherwise. And the number itself, forecast QTY, is not a fact about next month. It's what the rolling mean or the subagent's model computed from the flags it was given. Same flags, same path, same number every time. Claude picks compute it yourself
Carry-out
[1:31]Or hand it to a subagent by the flags, not by guesswork. And whichever path runs, the number it returns is a computed estimate, not a fact about next month. Your turn. Paste this into
Your turn
[1:41]Claude. I'm forecasting demand for product with a promotion next month, a 30-day horizon. Should you compute this yourself with a simple average or delegate it, and why? Then walk me through what confidence score you attach, and whether that's high enough to auto order or should escalate to a human instead. That's the actual test of a forecasting design.
Outro
[1:59]Flags decide the path. Liam in for bear.
