🤖 QA + AI

How to fit Claude and other LLMs into a real QA process — no magic, no hype. The model doesn't replace the tester: it removes the busywork so you can spend your head on risks and scenarios that actually need it. Below: articles and handy cheatsheets.

Articles

Cheatsheet: the prompt formula

Role+ Context+ Task+ Format+ Constraints

"You're a senior mobile QA" → "banking app, card-payment form, RU region" → "give negative cases" → "table: steps / result" → "poor network only, no duplicates".

What Claude closes — and what it doesn't

✅ Closes well

  • Test cases and checklists from requirements or a screenshot
  • Log and stacktrace analysis — find and group errors
  • Test-plan and requirements review — gaps and contradictions
  • Draft bug reports from raw notes
  • Test data: boundary, invalid, edge cases
  • Automation skeletons and manual-to-code conversion
  • Routine writing: release notes, emails, translations

⚠️ Don't trust

  • Your system's real behavior — it guesses
  • Fresh versions, limits, APIs — verify them
  • Counts and math without data
  • The final call on quality — that stays with you
  • Secrets, tokens, user PII — never paste them

Cheatsheet: quick templates

Test cases from a feature
You're a senior QA. Feature: [description]. Constraints: [platform, region, limits]. Generate test cases grouped by: positive / negative / boundary / security / network. Format — a table: # / steps / expected result. No duplicates.
Log analysis
Here's a log [paste]. Find the root error, group the repeats, explain unfamiliar exceptions in plain words, and suggest what to check first.
Requirements review
Read these requirements [paste]. Find contradictions, ambiguity, invalid assumptions, missing scenarios (errors, offline, permissions). Ask questions where unclear.
Bug report from notes
Turn notes into a bug report: [notes]. Format: title / steps / actual / expected / environment / severity. Flag what is missing.
Test data
Give test data for the field [type: date / money / name]: valid, boundary, invalid, attacks (injections, unicode, emoji), local formats for [region].
Case → automated test
Convert a manual case into a Playwright test with Page Object: [case]. Selectors by data-testid, explicit waits, no hardcoded delays. Point out where it is fragile.

Three techniques that raise quality

More deep-dives and practice on the Telegram channel.