Part 3 of 10 · 2 min read
Why a design answer needs a fixed shape, not just a diagram
A five-minute sketch almost always covers the same handful of areas and misses the rest.
Think of a building inspector who only ever checks the front door. A tour that never gets past the entrance will miss the wiring, the plumbing and the roof. It misses them no matter how carefully the entrance itself was examined.
In plain words: a typical quick system-design answer covers requirements, a queue, and maybe cost and streaming, but almost always leaves out durability, isolation, versioning, safety and operations — exactly the areas that distinguish a production design from a sketch.
A worked example. Checking a five-minute first draft against a 14-item rubric (requirements and SLOs, async job API, queue and workers, durable execution, streaming, cost controls, reliability, tenant isolation, deploy and rollback, versioning, observability and evals, safety and approvals, data and retention, incident response) finds it covers only 5 of 14 items — it states requirements, picks a queue, a streaming choice and a deploy choice, but says nothing about durability, isolation, versioning, safety, evals, or incident response.
The key lines.
# from code/mini-design/design.py
def review(design):
"""design: dict mapping rubric keys to a non-empty string saying what you chose. Returns the missing keys."""
return [k for k in REQUIRED if not str(design.get(k, "")).strip()]
Read it as: the rubric is just a checklist of keys, and a draft "passes" a key only by actually saying something for it. That is why running a real draft against it exposes the gaps a diagram alone would hide. Two of the choices that tend to get skipped in that missing nine are easy to name. One is the agent's autonomy level, meaning how much it acts without a human, from suggestions only to fully autonomous. The other is whether any step should fan out into several parallel sub-agents.
SLO, SLA, autonomy level, fan-out
An SLO is an internal target (such as "p95 resolution under 2 minutes"); an SLA is a looser contractual promise. Autonomy level is how much an agent acts without a human; fan-out is running several sub-agents or calls in parallel for one task.
Example: the 9 missing rubric items above, several of which are exactly decisions about autonomy, isolation and operations that a five-minute diagram skips past.