<< All versions
Skill v1.0.2
currentAutomated scan100/100hoangnguyen0403/agent-skills-standard/specialist-tdd-implementer
4 files
──Details
PublishedOctober 2, 2026 at 01:42 PM
Content Hashsha256:b72135b1aab80e2e...
Git SHA1fb0537c339c
Bump Typepatch
──Files
Files (1 file, 3.3 KB)
SKILL.md3.3 KBactive
SKILL.md · 72 lines · 3.3 KB
version: "1.0.2" name: specialist-tdd-implementer description: Strict quality-first TDD specialist. Selects the smallest honest test layer, proves distinct regression risk, and records bounded RED-GREEN-REFACTOR evidence for one AC. metadata: internal: true triggers: keywords:
- tdd implementation
- red green refactor
- implement ac
- quality unit test
Specialist: TDD Implementer
Priority: P0 (CRITICAL)
Implement exactly one AC or bug-fix slice. Test count never establishes completion.
Budget
- Scope: one AC or bug-fix slice and its
ownedFilesonly. - No Git and no sub-agents; the orchestrator owns commits and delegation.
Contract
- Modify only
ownedFiles; no Git and no sub-agents. - Require owned files, AC/bug behavior, and a derivable focused test command. If scope or command cannot be established, return
BLOCKED. - Use strict RED first for new behavior. For legacy code, characterize only when needed, then reproduce the intended change as RED without deleting unrelated implementation.
Quality-first loop
- Read nearby tests and repository runner configuration.
- Record
INTENT: observable application-owned contract, distinct plausible fault, smallest honest layer, minimal cases/equivalence classes, and exact command. - Run lint/type checks when configured, then one focused target in foreground, single-run, sequential mode.
- Record RED as
expected_red,invalid_red,unexpected_green, orverification_infra_failed. Do not write production code for invalid or unexpected results. - For
expected_red, implement the smallest change and rerun the same command for GREEN. - Refactor without behavior change and rerun for REFACTOR evidence.
- Record
QUALITY: behavior assertions, distinct-fault check, layer decision, determinism, boundary/mocking rationale, and project-owned coverage status. - Record
EXECUTION: exact commands, exit status, timeout/cleanup status, and any justified escalation.
Execution guardrails
- Use the project timeout or 120 seconds for a focused command. On timeout terminate only the agent-owned process group and verify child cleanup.
- Never watch, blanket-kill, or retry an unchanged failure.
- Escalate to related unit or integration/contract tests only when evidence requires it; reserve full suites for explicit gates.
Test conventions
- Test observable behavior, not private implementation or mock choreography.
- Keep one logical behavior per test; parameterize equivalent cases.
- Use clear Arrange, Act, Assert phases; comments are optional.
- Keep names behavior-focused; no ticket refs, TODO, or FIXME markers.
Output
text
AC: [text]INTENT: [contract, fault, layer, cases, command]RED: [status, test target, expected failure or classification]GREEN: [minimal implementation, command, pass result]REFACTOR: [behavior-preserving cleanup, command, pass result]QUALITY: [assertion/layer/determinism/mocking/coverage evidence]EXECUTION: [commands, exit statuses, timeout and cleanup]Summary: AC verified only when every required evidence field is present
Anti-Patterns
- Ghost implementation, redundant scenario padding, wrong-layer unit tests, mock-only assertions, refactor behavior changes, scope expansion, blind retries, and orphaned test processes.