Department 04 Test · Triage · Truth

QA.

QA with Claude isn't "AI finds the bugs for you." It's sharper than that: Claude executes the repro, decodes the crash, drafts the report, and expands the automated suite — so your time goes to the thing machines can't do, which is knowing where the game is going to break next.

00 — WHAT CHANGES

Three shifts in your day.

A

Repros run themselves

Through the Unreal MCP, Claude can attempt a reported repro in the editor itself — load the map, perform the steps, capture the state. "Could not reproduce" gets a lot rarer.

B

Crashes arrive pre-triaged

Every crash log comes back decoded: what failed, where, whether it's a known issue, and which change likely introduced it — before an engineer ever gets pinged.

C

Your reports become gold-standard

Claude turns raw session notes into structured, engineer-ready bugs — versioned, symbolicated, deduped against the existing database. Ping-pong dies.

01 — YOUR SKILLS

The playbooks you'll use most.

Browse all QA skills in the Library →

automation-testsRun, read, and grow the automated suiteskill
What it does

Runs Unreal's Automation framework with our filters — full suite or one system — and reads the results back in plain English. Better: hand it a bug you just caught manually, and it writes the regression test that makes sure it never comes back.

The suite grows a little with every bug you find. That's how manual QA compounds into automated coverage.

Try it
you › Run the full automation suite on the current build and give me a plain-English summary: what failed, what's flaky, what's new since Friday.
you › I found a bug where dropping an equipped weapon during a dash dupes it. Write a regression test for that exact sequence.
log-and-crash-triageCrash logs into plain Englishskill
What it does

Paste any crash, ensure, or log excerpt. Claude symbolicates it, identifies the failing system, checks it against our known-issues table, and tells you whether it's new, a dupe, or a regression — plus which recent changes are the likely suspects.

You stop being a courier between the crash folder and the engineering channel. You send diagnoses, not dumps.

Try it
you › Three crashes from tonight's session [paste]. Which are dupes of each other, which are known, and what do I file for whatever's new?
verify-in-editorClaude attempts the repro itselfskill + mcp
What it does

Give Claude the repro steps and it drives the editor: loads the map, performs the sequence, queries actor state, captures logs and screenshots at the moment of failure. It can also re-verify a fix the moment engineering lands it — closing the loop the same day.

Try it
you › Bug #2231 says the elevator soft-locks if you jump during the door close. Try the repro in Chapter1_Elevator, capture the state when it locks, and attach everything to the ticket draft.
you › Engineering says #2231 is fixed in this morning's build. Re-run the repro and confirm.
build-and-cookBuild literacy for the QA deskskill
What it does

Which build has which changes, why the nightly failed, whether the fix you're waiting on is actually in the package you're testing. Claude reads build logs and changelists so "what build am I even on?" always has a fast, accurate answer.

Try it
you › Is the fix for the vendor duplication bug in build 4312? If not, which build will have it?
release-checklistNothing ships with a step skippedskill
What it does

Our full pre-release gauntlet — packaging settings, platform requirements, cert checks, smoke passes — as a checklist Claude walks and evidences item by item. Release candidates get boringly predictable, which is exactly what you want.

Try it
you › Run the release checklist against RC3. Give me a table: item, status, evidence, and anything that needs a human decision.
02 — BEYOND THE BUILD

The rest of the job gets faster too.

01

Reports that write themselves

Raw session notes in, structured bugs out — repro steps, expected vs actual, build number, severity rationale, deduped against what's already filed.

02

Test plans from specs

Hand Claude a feature spec and get a draft test plan back: happy paths, edge cases, and the nasty interaction cases humans forget until Friday at 6pm.

03

Trend spotting

"What are our flakiest tests this month? Which systems generate the most regressions?" Claude reads across the bug database and tells you where the fire actually is.

Feed the system

Every mystery you solve becomes a lookup.

Cracked a weird failure that took two days? Add it to log-and-crash-triage's known-issues table with one sentence: "add this crash signature and fix to the skill." The next time it appears, triage takes ninety seconds — for anyone on the team. See exactly how →

← Prev dept
Art & Animation