Complete Slice

Take one slice from SLICES.md and ship it. Often for dev work, but might be a different task that you should still complete to the standards of the company

Process

1. Read the slice and the brief. Find the target slice. Internalize what it delivers. Read PROJECT_BRIEF.md for architectural context

1. Orient in the codebase. Module boundaries, test patterns, naming, DI style, config approach, what prior slices built. Your code should feel native — same seams, same style. Greenfield: scaffolding + test runner + lint + CI belong inside slice 1, not as separate phases

1. Plan the TDD sequence. Map acceptance criteria to red-green-refactor cycles. Start with the criterion that proves the core approach. Decide what's a real collaborator and what's a system boundary to mock. Keep the plan light. Use judgement for when to use TDD or take a different approach

1. Build one test at a time. Red → green → refactor. One failing test, minimum code to pass, clean up while green, next. Public interfaces only; mock system boundaries only; tests survive internal refactors. Writing all tests first produces tests that describe imagined behavior, not actual. See [tdd-reference.md](tdd-reference.md) for the full methodology

1. Bake SRE in as you go. Structured logs at decision points and error paths; metrics on the behaviors this slice introduces; health checks on new endpoints; errors that fail explicitly at boundaries with context; config externalized. Operational readiness isn't a final step — if it's not present when tests go green, it won't be added

1. Verify. All acceptance criteria met and covered by passing tests. Observable outcomes demonstrably work — run it, hit the endpoint, read the logs. Full suite green, not just new tests. No unrelated changes mixed in

1. Consider future agents by running updating-ai-knowledge, often for AGENTS.md

1. Close the slice. Check the boxes in SLICES.md. Follow-up work, discovered edge cases, or risks land in the relevant future slice, or in a ## Notes section at the bottom of SLICES.md

1. Report what you shipped and verified back to whoever ran you

Scope

Feature work, infrastructure-as-code, migrations, instrumentation, perf, bug fixes, deploy automation. Language, framework, and cloud agnostic

Not for: research spikes with no code deliverable, pure documentation, project planning (use plan-to-slices)

Skills you lean on: the verify subagent before declaring done, and a subagent for a second opinion or implementation partner

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ryan-scheinberg
• Source: ryan-scheinberg/harness
• License: MIT

Install and usage instructions live in the source repository linked above.