Generate Principal-Engineer-grade system design interview preparation documents. Use when the user says "design a system", "system design for [company]", "prepare [system] for PE interview", "break down [system]", "system design interview", or names any system design problem (URL shortener, distributed cache, news feed, rate limiter, top-K, chat system, etc.). Produces 11 artifacts per problem (8 design docs + 1 Excalidraw diagram + 1 eval report with PASS/FAIL evidence + 1 independent interview transcript) covering requirements, NFRs, entities, API, schema, HLD, deep dives, bottlenecks/tradeoffs, adversarial technical-depth review, and machine-verified quality validation. Calibrated for Principal Engineer interviews at Databricks, Anthropic, OpenAI, Google, Amazon, and Microsoft.
Generate Principal-Engineer-grade system design interview preparation documents. Use when the user says "design a system", "system design for [company]", "prepare [system] for PE interview", "break down [system]", "system design interview", or names any system design problem (URL shortener, distributed cache, news feed, rate limiter, top-K, chat system, etc.). Produces 11 artifacts per problem (8 design docs + 1 Excalidraw diagram + 1 eval report with PASS/FAIL evidence + 1 independent interview transcript) covering requirements, NFRs, entities, API, schema, HLD, deep dives, bottlenecks/tradeoffs, adversarial technical-depth review, and machine-verified quality validation. Calibrated for Principal Engineer interviews at Databricks, Anthropic, OpenAI, Google, Amazon, and Microsoft.
System Design Interview -- Principal Engineer Grade
Generate rigorous system design documents by reasoning through the problem, not
by recalling memorized architectures. Every design is a hypothesis that holds
until a constraint changes.
How to use this skill
Step 1 -- Read the orchestrator:
Read references/orchestrator.md (relative to this skill directory). It defines:
The 6-phase generate-evaluate-fix loop with per-phase quality gates (27
criteria across 6 gates)
Cross-file consistency protocol (5 checks across all output files)
Outer eval loop with PE rubric scoring
Validator harness that produces machine-verified 09-eval-report.md
Interviewer-Research loop that produces 10-interview-transcript.md
Context management protocol to prevent context overload
Just-in-time reference loading schedule
Step 2 -- Execute the orchestrator loop.
Follow the orchestrator exactly. It tells you which reference files to load at
each phase, when to create a context checkpoint, and how to run the inner
generate-evaluate-fix cycle (max 2 iterations per phase gate).
Step 3 -- Run the Interviewer-Research loop.
After Phase 4, Phase 5, and Phase 6, run the Interviewer checkpoint from
references/interviewer-protocol.md. The Interviewer reviews blind to
self-scores and flags depth gaps, fake bottlenecks, wrong technology choices,
missing failures, and scale holes. Run references/research-protocol.md only
for Major or Critical findings; minor findings are fixed locally unless
they repeat. Append the critique, conditional Research findings, and revision
log to 10-interview-transcript.md.
When the platform supports model selection, prefer a different model family for
the Interviewer to reduce self-eval leniency. Otherwise, use the same model with
the Interviewer protocol.
Step 4 -- Run the outer eval loop.
After all 8 design files pass their phase gates and the Interviewer-Research
loop has closed major/critical findings, score against the PE rubric using the
Interviewer findings as primary evidence. If below bar (avg < 4.5 or any
dimension < 4), revise the weakest file section. Max 1 revision pass.
Step 5 -- Run the validator harness.
After the outer eval loop completes, run the Python validator. Find the
validator at scripts/validator/ relative to this skill directory. Run it
against the output folder:
python3 -m validator validate <output-folder>
Run this command from the scripts/ directory within this skill, or use an
absolute path. This produces 09-eval-report.md with PASS/FAIL evidence for
every gate criterion, cross-file consistency check, quality signal, and
technical-depth check. If FAIL, fix flagged issues (max 1 pass) and re-run.
Output location: Generate all files in system-design/<problem-name>/
relative to the user's project root (the repo where the user invoked the
skill), not relative to this skill's install location.
The Principal Engineer bar (quick calibration)
PE is not "Staff but more." It is a qualitatively different signal:
Reframes before solving. Questions the problem itself. "Do users need
exact ranking or would approximate trending suffice? That changes the entire
architecture."
Sees the full solution space. Lays out 2-3 architectures, explains the
forces that pick between them, selects with judgment. Pivots instantly when a
constraint changes.
Brings production war stories. Not "Redis can do this" but "We ran Redis
at 2M QPS and the failure mode that bit us was..."
Thinks organizationally. Team ownership, blast radius of a bad deploy,
oncall burden, cross-team dependencies.
Anticipates the 3-5 year arc. What breaks at 10x, 100x, 1000x. What to
rebuild vs evolve.
Teaches the interviewer something. Novel insight, non-obvious tradeoff, a
connection between subsystems that surprises.
YAGNI at scale. Chooses the cheapest design that satisfies constraints and
explains why the fancier option is not justified yet.
Hard rules
No name-dropping. Every technology choice gets the trade-off triad:
solves / worsens / when-to-change.
No design before numbers. Capacity estimation precedes architecture.
No vague boxes. If you cannot write the request, response, and primary
key, the component is guesswork.
No infinite uptime. Every component fails. State the degradation story.
"More retries" amplifies outages.
No defending sunk designs. When a constraint changes, name the
invalidated assumption and redesign only the affected part.
No complexity for its own sake. The cheapest design that meets the
constraints wins. Justify every added component.
excalidraw-diagram -- Used at Phase 4b (after Gate 4 passes) to generate
a visual .excalidraw companion diagram for the HLD. Bundled alongside this
skill in the same repository.
Exemplars
Use assets/exemplars/in-memory-cache (exemplar for NFRs, latency budgets,
runbooks) or assets/exemplars/trending-articles-top-k (exemplar for deep
dives, capacity estimation) for output calibration. Read only the specific
file matching the current phase, not the full 8-file set.
