skills/implement/SKILL.md
Atomic TDD build skill. Takes a context packet (shaped ticket) and produces code + tests on a feature branch. Red → Green → Refactor. Does not shape, review, QA, or ship — single concern: spec to green tests. Use when: "implement this spec", "build this", "TDD this", "code this up", "write the code for this ticket", after /shape has produced a context packet. Trigger: /implement, /build (alias).
npx skillsauth add phrazzld/agent-skills implementInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Spec in, green tests out. One packet, one feature branch, one concern.
Delegation floor applies: probe the roster first; dispatch two or more
providers for substantive work; direct solo only for mechanical, emergency,
user-forbidden, or fewer-than-two-providers cases. See
harnesses/shared/AGENTS.md (Roster).
Local lane guidance: Use one bounded builder lane and one adversarial validator/refactor lane; give them the context packet, oracle, boundaries, repo anchors, and what would embarrass us if green tests missed it.
Completion evidence core applies: use harnesses/shared/AGENTS.md
(Completion Evidence) as the universal evidence shape, then fill the local
fields required by the implementation final message.
Input. A context packet: goal, non-goals, constraints, repo anchors, oracle (executable preferred), implementation sequence. Resolution order:
/implement backlog.d/033-foo.md)/implement 033) → resolves via backlog.d/<id>-*.md/shape output in the current sessionRequired packet fields (hard gate — missing any = stop):
goal (one sentence, testable)oracle (how we know it's done, ideally executable commands)implementation sequence (ordered steps, or explicit "single chunk")If the packet contains Formal Spec Required: yes, also require a Formal Spec block with: informal spec, formal examples, acceptance oracle, hardening
budget, and waiver path. Missing or vague formal-spec fields are packet
incompleteness, not builder judgment.
See references/context-packet.md for the full shape.
Output.
<type>/<slug> from current branch)Stops at: green tests + clean tree. Does not run /code-review,
/qa, /ci, or open a PR.
Resolve the packet (order above). Parse required fields. If any are missing or vague ("add feature X" with no oracle), stop with:
Packet incomplete: missing <field>. Run /shape first.
Do not try to fill in the gaps. Shape is a different skill's judgment.
git checkout -b <type>/<slug> from the current branch. Builders never
commit to master/main. If you forget, create the branch after and
cherry-pick before handing off.
Spawn roster-backed builder lanes with:
Builder prompt must include:
You MUST write a failing test before production code. RED → GREEN → REFACTOR → COMMIT. Exceptions: config files, generated code, UI layout. Document any skipped-TDD step inline in the commit message.
For packets with Formal Spec Required: yes, the first red test must be an
acceptance test derived from the formal examples and acceptance oracle. Unit
tests come after the failing acceptance check exists; production code comes
after both acceptance and unit intent are executable. Waive this only through
the packet's named waiver path and record the waiver in the Completion Gate.
See references/tdd-loop.md for the full cycle and skip rules.
Milestone critic gate. After each chunk's builder returns, dispatch a fresh
read-only critic that sees ONLY the chunk diff + the packet oracle + the todo —
never the builder's reasoning. It returns blocking gaps (oracle clause unmet,
behavior lost, invariant violated). Do not advance to the next chunk until the
critic returns no blocking gap, or the gap is explicitly waived in the
Completion Gate. Skip for trivial diffs (<20 LOC, single file). This enforces
the milestone gate defined in harnesses/shared/AGENTS.md (Layer 2 — Dispatch).
Before exiting, confirm:
git status clean (no untracked debug files)TODO/FIXME/console.log added that isn't in the spec/hardening property, or the handoff names why example tests are enoughIf any check fails, dispatch a builder sub-agent to fix. Max 2 fix loops, then escalate.
Output: feature branch name, commit list, oracle checklist (which commands pass), Completion Gate, residual risks. Do not run review, do not merge, do not push unless the packet explicitly says so.
## Completion Gate
- Exact end-user behavior changed: behavior or internal operator behavior implemented.
- Evidence that proves it: failing-then-passing test, oracle command, or artifact proving the behavior.
- Exact command/path/route exercised: command, URL, route, file path, or tool call actually run.
- Repo-fit check: live repo pattern, contract, or boundary the implementation follows.
- Observability / instrumentation debt: named signal added, existing signal used, or debt recorded.
- Residual risk: unverified path, deferred edge case, or none with reason.
Completion evidence core applies: behavior changed or verified, live evidence,
exact command/path/route, repo-fit check, and residual risk. See
harnesses/shared/AGENTS.md (Completion Evidence).
Local fields include observability / instrumentation debt when the changed behavior has no named signal.
For internal-only changes, replace "end-user" with the developer/operator behavior the implementation changes.
/refactor's job, not yours./deliver / /flywheel)/shape)/code-review)/qa)/ci)/refactor)/deliver --polish-only → /ship)Stop with a loud report if:
Not stopping conditions: spec is hard, unfamiliar codebase, initial tests red. Those are the job.
console.log, print("here"), commented-out
code. The tree must be clean before exit.git checkout -b first.development
Lightweight evidence-backed retro and catch-up reports for a current repo, branch, PR, backlog slice, or recent agent session. Use when the user asks for a debrief, catch me up, what changed, why it matters, product implications, end-user implications, developer experience implications, current app state, backlog state, workspace state, alternatives considered, or context rebuild after losing the thread. Trigger: /debrief.
testing
Capture agent-session work records as local JSONL audit evidence. Links a backlog/spec, branch, commits, review verdicts, QA/demo evidence, transcript refs, and shipped ref without storing raw private transcripts. Use when: "trace this work", "write work record", "agent session trace", "journal this delivery", "link transcript evidence". Trigger: /trace, /journal.
data-ai
Turn proven agent-session patterns into first-party Harness Kit skills. Use when: "skillify this conversation", "make this into a skill", "generate a skill from current transcript", "extract reusable workflow". Trigger: /skillify.
testing
Run one targeted, read-only architecture or quality critique through a named lens from the shared rubric. Use when: "critique this module", "run an Ousterhout pass", "lens critique", "architecture critique". Trigger: /critique.