Cognitive Drills

🧠 Cognitive Drills: train the engineer’s thought process

Why

Knowledge is cheap (AI can recite it). What gets you hired and promoted is how you think: predicting behavior, reasoning under uncertainty, finding root causes, and making trade-offs explicit. These drills train that deliberately.

The 14 drills

#	Drill	How	Example
1	Predict → Verify ⭐	Before any experiment, write a prediction with numbers. Run it. Explain the gap	”Halving the Hikari pool from 20 → 10 will raise p99 from 40 ms to ~?“
2	Feynman teach-back	Explain a concept in 3 min, out loud, no notes; record it; note where you stumbled	”Explain MVCC to a junior”
3	Blank-page design	Design a system from memory on paper in 20 min, then compare with references	”Design Kafka”
4	Constraint flip	Take a finished design and change one constraint: 100× traffic, strong consistency, ½ budget, multi-region, 10 ms latency	”Orbit engine with 1M runs/s?“
5	Pre-mortem	”It’s 6 months later and this failed badly. Why?” List 10 causes, rank them, mitigate the top 3	Before each Orbit version ships
6	Symptom → hypotheses	Given symptoms only, write 3–5 ranked hypotheses + the cheapest test for each	”p99 spikes every 30 s, CPU flat”
7	Reverse engineering	Read a product’s behavior/API and infer its architecture before reading its engineering blog	Discord message storage, Stripe idempotency
8	Trade-off debate	Write the best argument for and against an option, then decide	”Kafka vs Postgres queue for tasks”
9	Fermi estimation	5-minute order-of-magnitude estimates with explicit assumptions	”Tokens/day for 10k tenants × 50 runs”
10	Code archaeology	Read unfamiliar source, draw the call graph, and explain one design decision	ThreadPoolExecutor.execute()
11	Compression	Summarize a chapter/paper in 5 bullets, then in 1 sentence	DDIA ch. 7
12	Transfer	Apply an idea from domain A to domain B	”Kafka’s HW idea → Orbit’s safe-to-archive marker”
13	First-principles derivation	Derive why a mechanism must exist from the failure it prevents	”Why does Raft need terms?“
14	Self-review after 7 days	Reread your ADR/code from last week as a critical reviewer; write 3 critiques	Every Sunday

Routine

• Daily (15 min): 1 Predict→Verify during Build time + 1 Fermi or teach-back
• Weekly (60 min, Saturday): Blank-page design + Constraint flip on that week’s topic
• Per Orbit version: Pre-mortem before, self-review after

Drill log

Date	Drill	Topic	Prediction / claim	Reality / critique	Lesson

Prompt bank: symptom → hypotheses (drill 6)

• p99 latency spikes every ~30 s; CPU and QPS flat
• Kafka consumer lag grows only for 1 of 12 partitions
• After a deploy, 1% of requests 502 for ~20 s
• Postgres CPU 90%, the same query count as yesterday
• Memory of a Go service grows 50 MB/hour, never drops
• A Java service with virtual threads has lower throughput than with platform threads
• The LLM gateway shows TTFT doubling at 3 pm daily
• RAG answers got worse after re-ingesting the same documents
• An agent occasionally performs the same tool action twice
• gRPC traffic all goes to one of 5 pods