Tld Skills — Testing skill for Claude Code

Adventure Skills

A set of Claude Code skills for **Test-Led Development (TLD)**. Drives a project through small Jira tickets one at a time, with hard stops between phases, drift detection against the ticket spec, side quests in isolated git worktrees, and milestone-boundary gate checks.

---

Philosophy

**Test-led, not test-driven.** In TDD, tests are a discipline you follow while writing code. Here, tests *are the specification*. If you can't write a test for an acceptance criterion, the criterion is underspecified — fix the ticket, not the code. The implementation is whatever minimum amount of code makes the tests pass.

**Hard stops between phases.** Each ticket runs the same sequence: **red → review → green → verify → commit**. Every transition is a mandatory stop. After failing tests are written, the framework pauses for you to review the spec before any implementation is written. After verification passes, it pauses again for manual QA approval before committing. Silence is not approval; only an explicit approval keyword advances the flow. This is the central guardrail — it stops the agent from racing ahead and shipping work you haven't actually inspected.

**Side quests in isolated worktrees.** When you notice a small unrelated fix mid-ticket, `/tld-side-quest` spins up a separate git worktree, handles the fix there, and returns you to the original ticket without polluting the main context or leaking unrelated changes into the active commit.

**Drift detection.** Before committing, the framework re-reads the ticket's acceptance criteria and checks the actual diff against them. If the implementation has drifted from the spec — extra files, missing AC items, scope creep — the gate flags it and asks before proceeding.

**Milestone boundaries are real gates.** After every ticket in a milestone is Done, `/tld-gate` runs the full regression suite, verifies every ticket's status, and confirms the milestone's exit criteria are met before the next milestone ope