04 · Drill 01 — 30-Second Model Selection

Status: Outline. Body fills in Week 2. Voice: principal-level, BFSI-threaded, Apic-calibrated.

What this drill is. Cold delivery of the Claude lineup decision matrix in 30 seconds — pick the right model for a given use case description, defended on capability + cost + latency. Ten use-case cards, randomized.

What this drill is NOT. A capability comparison drill. Not benchmark recital.

Prompt

Partner picks one of the 10 use-case cards (below) and reads it aloud. You have 30 seconds to: (a) name the model, (b) name the tier (Haiku / Sonnet / Opus), (c) name the one trade-off you accepted.

Time box

Per card: 30 sec.
Drill: all 10 cards = ~6 min.
Hard cap: 35 sec — beyond is Lean No on this drill (the matrix isn't reflex).

The 10 use-case cards

"Triage classifier in front of a support agent assist — 200K calls/day, sub-200ms latency budget, 5-class output."
"Compliance document review for a public-sector regulator submission. Failure = retract submission."
"RM copilot at a wealth-management desk — multi-turn agentic, customer 360, occasional plan-and-execute."
"Internal SOP search RAG over a 50K-document corpus. ~10K queries/day."
"LLM-as-judge for an eval harness scoring faithfulness on 5K candidate responses per run."
"Exec analytics natural-language interface over Snowflake — 100 queries/day from C-suite users."
"Developer productivity copilot inside an internal IDE plugin — streaming, sub-3s latency budget."
"Customer-facing chatbot on a public site for product Q&A — 50K calls/day."
"Multi-document compliance gap analysis across 200-page regulatory bundles. Run weekly, ~100 runs."
"Tool-use planner that decides which of 12 tools to call on a Salesforce-integrated workflow."

Strong Hire answers (placeholder — fill in Week 2)

Card	Answer	Trade-off accepted
1	Haiku 4.5, hot path	Capability ceiling — escalate complex cases to Sonnet
2	Opus 4.7 with eval gates	Cost — but the failure mode justifies it
3	Sonnet 4.6 default, Opus on hard	Two-model stack adds eval complexity
4	Sonnet 4.6 with caching	Cache pollution risk on doc updates — design for it
5	Haiku 4.5 as judge	Judge calibration must be tighter on a smaller model
6	Sonnet 4.6 + Haiku for query routing	Latency on the routing hop
7	Sonnet 4.6 streaming	Cost vs Haiku — but capability wins for code
8	Sonnet 4.6 with caching	Per-call cost; volume justifies caching investment
9	Opus 4.7 batch	Latency — but it's a batch run
10	Sonnet 4.6 for the planner, called tools per their needs	Planner-vs-executor split adds an eval layer

Rubric

Strong Hire

Model named in <10 sec.
One specific trade-off named (not "it's a balance").
Tier-architecture pattern surfaces when the use case justifies it.
No "always Opus" or "always Haiku" defaults.

Hire

Model named, but trade-off is generic.
Two-model stack mentioned only when prompted.

Lean No

"It depends" without the deciding axis named.
Model picked without naming a trade-off.

Strong No

Wrong model ID (e.g. "claude-3-opus" — old).
Single-model hammer for every card.

Common Lean No traps

Trap 1 — Always picking Opus

Conservative, but reads as "I haven't done the math."

Trap 2 — Marketing-speak

"Sonnet is a great workhorse" — without the why (capability fit, cost shape, latency).

Trap 3 — Skipping the trade-off

The rubric requires a trade-off named. Skipping it = Lean No.

Trap 4 — Wrong/stale model IDs

Using claude-3-opus, claude-2, etc. Strong No.

Trap 5 — Ignoring the latency or volume signal

Card 1 says "sub-200ms latency, 200K calls/day." Picking Sonnet ignores both signals.

How to run this drill

Partner shuffles cards. You get them cold.
Timer starts when card is read.
30 sec to answer. You get scored on model + trade-off.
Log to Drill Tracker.
Application trigger: 9 of 10 cards at Strong Hire across 2 sessions.

Cross-references

Strong-Hire bar for this drill

9/10 cards delivered in <30 sec.
Trade-off named on every card.
No stale model IDs.
Tier-architecture pattern reflex when justified.