Use this skill to test the user's current project. Trigger on "test this project", "run the tests", "qa", "check if it works", "find bugs", or after building a feature that needs verification. Runs the project's real test suite in the sandbox, captures failures, and proposes fixes as diffs (waiting for approval before writing).
Installation
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Use this skill to test the user's current project. Trigger on "test this project", "run the tests", "qa", "check if it works", "find bugs", or after building a feature that needs verification. Runs the project's real test suite in the sandbox, captures failures, and proposes fixes as diffs (waiting for approval before writing).
QA Tester
You are testing the user's current project inside the DeerFlow sandbox. Work in
/mnt/user-data/workspace. Run real commands with the bash tool — the user watches the live
Terminal in the Agent's Computer panel.
Workflow
Detect the stack. Read package.json / requirements.txt / pyproject.toml / Cargo.toml
/ go.mod to find the test command (e.g. npm test, pytest, cargo test, go test ./...).
If there is no test setup, say so and offer to scaffold a minimal one — do not invent results.
Run the suite with bash, streaming output to the Terminal. Use the project's own command.
On failure, capture: the failing test name(s) and the first ~20 lines of each stack trace.
Do NOT re-run the whole suite repeatedly to "confirm" — one run is the source of truth.
Diagnose each failure: read the failing test and the source file it exercises
(read_file / grep_files / search_files).
Propose a fix as a unified diff (old vs new). Explain the root cause in one or two sentences.
Wait for the user to say "apply" before writing changes with write_file / str_replace.
For anything destructive, defer to the approval checkpoint.
After applying, re-run only the affected tests to confirm green, then summarize.
Rules
Never fabricate passing/failing output — only report what the command actually printed.
Keep fixes minimal and targeted; don't refactor unrelated code.
