A daily training rig for product judgment, built as two Claude skills that work as a pair.
judgment-drill runs a 25-minute session across five stations, every case sourced fresh from the web with a publicly checkable number behind it. judgment-coach runs after it and works out why the answers failed — which reasoning move produced each miss, whether that move recurs, what procedure defeats it, and where the same error is currently costing someone else money.
The design principle behind all of it: commit to an answer before seeing the ground truth. Experience accumulates on its own. Judgment only accumulates when a prediction gets scored.
Most "practice" for senior product work is reading case studies, which trains recognition rather than judgment. You read what happened, it makes sense, you file it away, and nothing about your future estimates or diagnoses changes. Nothing was ever at stake because no prediction was ever made.
This rig makes you commit first. Every station takes an answer and a stated confidence before revealing anything, then scores both. Correctness tells you whether you were right today; calibration tells you whether your confidence can be trusted tomorrow, and the second is the actual deliverable. A leader who is right 70% of the time and knows it is worth more than one who is right 80% of the time and claims 95%.
The coach exists because a score is not a diagnosis. Being told an answer was worth 2 of 3 doesn't tell you which reasoning move produced the gap, whether you have made that move in four of the last six sessions, or what a strong answer's procedure actually looks like.
| # | Station | Trains | Box |
|---|---|---|---|
| 1 | Fermi estimate | Reasoning to a number from priors | 4 min |
| 2 | Stock or flow | Diagnosing the shape of a change | 4 min |
| 3 | Third-order effects | Consequence chains at scale, after the affected party adapts | 6 min |
| 4 | Reverse the memo | Reconstructing someone else's strategic logic in 120 words | 8 min |
| 5 | Go problem | Sente/gote, urgency, reading a group's life before investing in it | 3 min |
Cases are pulled fresh every session and must clear a verification bar: a real number a real person could open a link and check. Small and obscure beats famous — an indie SaaS with a public revenue dashboard teaches more than a headline case you already have opinions about.
The Go station is not there to make you strong at Go. It installs vocabulary — sente, gote, urgent-before-big, thickness, dead group — for moves in product strategy that otherwise stay unnamed. Once a decision has a name, it can be argued about.
After grading, the coach runs four passes:
-
Classify — every miss gets a code from a fixed taxonomy of ~25 reasoning failures (
C1 SINGLE-CAUSE-CAPTURE,C3 WRONG-METER,S5 DURESS-MISSED,M4 RIGHT-FOR-WRONG-REASON…). Each diagnosis must quote your words, name the move upstream of the answer, and point at what you walked past. Codes stay stable across months, which is the only reason recurrence counts mean anything. -
Root — codes get merged into the one or two structural weaknesses they share. Most people have two real weaknesses wearing many costumes; reporting eight separate errors hides that.
-
Method repair — the procedure that defeats each code, run against the case you just failed, showing the answer it produces. All three parts, every time: steps alone are platitudes, a worked example alone doesn't generalise.
-
Casebook — two to four real, sourced cases where the same move (not the same industry) cost someone something documented.
Then it writes a diagnosis line to the shared log with a target: for the next session — a case property the drill sources against, so tomorrow's cases aim at today's weakness instead of being random.
The highest-priority code in the taxonomy is M4 RIGHT-FOR-WRONG-REASON: a good score reached by a method that would fail on a slightly different case. The drill's grading can never surface it, and you'll keep running the method until it breaks on something expensive.
/plugin marketplace add orchimada/judgment-drill
/plugin install judgment-drill@judgment-drill
Both skills install together. Replace the owner with your fork if you're running a modified copy.
Copy the two skill folders into your skills directory:
cp -r skills/judgment-drill ~/.claude/skills/
cp -r skills/judgment-coach ~/.claude/skills/Works in Claude Code and in Cowork. The drill needs web search to source cases; both skills work best with a persistent memory filesystem, and fall back to a local judgment-drill-log.md without one.
/drill full session
/drill 1 single station (or "fermi", "stock or flow", "third order",
"reverse the memo", "go")
/drill review weekly review — calibration curve, trend by station
/drill log recent history
/coach full diagnosis on the session just graded (runs automatically)
/coach pattern cross-session pattern only, no new cases
/coach method 3 the attack angle for one station, cold — useful *before* a session
/coach casebook C3 fresh real-world cases for one failure mode
Plain language works too: "daily drill", "quiz me", "why do I keep getting this wrong".
Both skills share one file. The drill appends an S<n> session line; the coach appends a C<n> diagnosis line with error codes and recurrence counts.
S12 | 2026-08-14 | L2 | 2/3/1/2/3 = 11/15 | conf 85→60 | cases: … | go: sente/gote | note: …
C12 | 2026-08-14 | codes: C3(4), S1(2) | root: visible over operative | repaired: … | target: …
The cases field enforces rotation — no company or Go concept repeats inside twenty sessions. The codes field is what turns coaching into training: an error seen once is noted, seen twice is hedged, seen three times in six sessions is stated flatly and becomes the thing you work on. Errors that go quiet for four sessions get marked repaired, because a log that only accumulates accusations stops being read.
See docs/example-session.md for a full worked session — cases, answers, grading and the coach's diagnosis.
Three levels. Promotion needs three consecutive sessions at 12+/15 with calibration inside band. Escalating on raw score alone produces a confident, badly-calibrated user, which is the exact failure mode the rig exists to prevent.
A few decisions that look arbitrary and aren't:
- The answer never appears in the same message as the question. If it does, the session degrades into reading, which is the thing this replaces.
- Confidence is collected separately at every station and bucketed. Correctness and calibration are reported apart, and when they diverge the calibration finding leads.
- The coach may not re-grade. Re-litigating the score spends the session's remaining attention on its least useful artifact. The coach's unit of analysis is the move, not the answer.
- No pattern claims before ~4 sessions. Diagnosing a "tendency" from one data point teaches the user to distrust the diagnosis, which costs more than the missed observation.
- Cases without a checkable link are discarded, not reshaped. A fabricated number silently corrupts the calibration log, which is the one asset that accumulates.
- Word caps are load-bearing. The 120-word memo cap is the compression exercise; the coach's length cap exists because a diagnosis long enough to feel comprehensive is long enough to go unread.
Contributions of new error codes and casebook cases are welcome — see CONTRIBUTING.md.
MIT — see LICENSE.