2 min lesson

A tiny repro harness beats prose

Answer this as if it were happening now: "AI tools are allowed in the debugging screen and the agent suggests a fix that looks plausible. What is the strongest move?" Say what supports your choice.

Layer: Client (editor/Electron)
Tell-tale signs: Crashes, UI freezes, settings or extension conflicts, version-specific
How to confirm: Bisect versions, disable extensions, check the dev console and local logs

Layer: Server / network
Tell-tale signs: Timeouts, 5xx, auth or region issues, fails for many users at once
How to confirm: HAR capture, status of the request, retry, compare across networks

Layer: Model / agent
Tell-tale signs: Hallucinated or oversized diffs, lost context, agent loops, nondeterministic output
How to confirm: Re-run with a different model, shrink the context, check whether it reproduces deterministically

Layer	Tell-tale signs	How to confirm
Client (editor/Electron)	Crashes, UI freezes, settings or extension conflicts, version-specific	Bisect versions, disable extensions, check the dev console and local logs
Server / network	Timeouts, 5xx, auth or region issues, fails for many users at once	HAR capture, status of the request, retry, compare across networks
Model / agent	Hallucinated or oversized diffs, lost context, agent loops, nondeterministic output	Re-run with a different model, shrink the context, check whether it reproduces deterministically

Distinguishing client vs server vs model is the core triage axis for an AI-native desktop app.

A tiny repro harness beats prose - it makes the bug deterministic and hands engineering something runnable

# minimal repro: does the oversized-diff bug depend on the model or the context size?
for model in gpt-class claude-class; do
  for ctx in small large; do
    run-agent --model "$model" --context "$ctx" --task fixtures/refactor.md \
      --record runs/$model-$ctx.json
  done
done
# then diff the recorded edits to see which variable flips the failure

Interview move

When AI tools are allowed, use them and narrate your judgment. Ask the agent to draft a repro script or summarize a log, then say what you accept and what you reject and why. Pasting raw model output into a bug report is the anti-signal here; rejecting a plausible-but-wrong suggestion out loud is the signal they want.

End with a clean handoff artifact

General-industry pattern for support and escalation screens: they reward clarity, speed and a tidy artifact at the end. Even if you don't fully root-cause in the time box, close with a written summary - repro steps, expected vs actual, evidence captured, suspected layer and the next test you'd run. That artifact is often what is actually being graded.

A tiny repro harness beats prose

Step 1 of 2