System design interview

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.

References (all relative to this skill directory)

• references/orchestrator.md -- The generate-evaluate-fix loop, phase gates, cross-file consistency, outer eval, validator harness. Read first, always.
• scripts/validator/ -- Python CLI that produces 09-eval-report.md with machine-verified PASS/FAIL results.
• references/principal-engineer-bar.md -- The 10-dimension PE rubric and self-scoring.
• references/interviewer-protocol.md -- Independent adversarial review protocol for technical depth.
• references/research-protocol.md -- Conditional Research agent protocol for major/critical gaps.
• references/reasoning-engine.md -- The 6-phase design loop, failure modes, coverage sweep, curveball protocol.
• references/building-blocks-index.md -- L0-L7 bottom-up component catalog with PE depth signals.
• references/company-profiles.md -- Interview format and PE expectations for 6 companies.
• references/problem-bank.md -- 30+ curated problems with company attribution and curveballs.
• references/tradeoff-framework.md -- The 3-question method and common axis tradeoffs.
• references/numbers-to-know.md -- Latency table, QPS tiers, powers of 2, estimation recipes.

Companion skills

• 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.