# Implement > Use when the slice portion is locked and its derived tests exist, to build all of the locked portion to green autonomously. Triggers at implement start and on every re-entry after a human turn. Do not use to declare green (slice-exit does that), and do not use it for an unexplained runtime defect (problem-solve). - Skill: `teklabsdigital/implement` (Agent Skill) - Install (CLI): `npx skillmds@latest add teklabsdigital/implement` - Raw SKILL.md: https://api.skillmd.com/api/skills/teklabsdigital/implement/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: teklabsdigital (https://skillmd.com/u/teklabsdigital) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/teklabsdigital/implement --- # X2 Implement Build the slice: generate, verify, correct, re-verify, running to green without human turns wherever you honestly can. The kernel enforces the invariants; this skill exists to capture the decisions the kernel cannot make and to keep the drive honest. ## The contract you are building to **All of the locked portion.** Fidelity to the whole approved contract, no silently dropped pieces (this is what gate 3 ratifies, through the exit report). The pilot's dark-only theme flip is the canonical violation: a visible piece of the locked contract quietly dropped while everything read as done. "All of it" means all of the locked portion; the rest of the prototype is the design spine awaiting its slices. ## The human-turn contract (stop only for these) - **Record-gap**: a decision no record answers (not in D-000, the deltas, or the locked prototype) that is novel and not cheaply reversible. File a one-line question; the answer becomes a D-0xx. Commodity decisions you make yourself from priors with a one-line note. A real example of a record-gap: a reuse-versus-build call on an existing component; the kernel is provenance-blind by design and the economics are local to the slice. - **Green-but-wrong**: tests green, result wrong against the story or the locked prototype. Surface it; the fix goes in the decision file, never patched into source. - **Unconditional turns**: any persistent-schema migration or published-contract change, every time, no exception. - **Suspected kernel defect**: stop and report it with why it looks like a kernel defect. The kernel is fixed canonically, never patched locally without a ruling. When unsure whether something is a turn, surface it and let intake classify it. Never absorb it silently, and never manufacture a turn for a commodity choice. ## Standing rules, now held by the gates These were settled by the record and, since the invariants pass (2026-07-11), are enforced by the kernel; the notes here say what the gate will do so you work with it, not around it. - **The client UI is a thin layer over testable client services** (UI-5). The composition root constructs the client and delegates; screens receive data and callbacks. The eslint import ban rejects transport or data-service imports outside the composition root (type-only imports are legal), and the composed-entrypoint smoke proves a real request leaves the app per primary flow. Each NEW primary flow owes its smoke line; that is your per-slice obligation, checked in the exit report's coverage table. - **Every external side effect sits behind a provider interface bound at the composition root** (delivery, model calls, clock-like effects). This is what let the pilot close a security-sensitive e2e gap as a binding-only change instead of a redesign (CONF-02); it is also what the gated harness profile needs, so an inline-instantiated effect is unreachable by e2e. Stated in the edition README; review each new external effect for it. - **Operational settings are configuration, not code** (CFG-1). Non-secret settings resolve from committed config, secrets from the secret store; the architecture test bans a literal registry (model ids, provider endpoints) in host source and scripts. When the product gains a new operational-setting shape, extend the registry at D-000 rather than letting it pass. - **Harness completeness is self-auditing** (TEST-2). Every public method of every client data service needs a scenario through the real transport, and the harness fails on an uncovered method, so write the scenario with the method, not after the interrogation. Where a scenario needs a seam the server does not expose, the ruled shape is the gated non-production harness profile: re-bind provider interfaces at the composition root under `Harness:Enabled`, which refuses to boot outside Development/Testing; never a weakening of a production path. ## The ledger Append every human turn to the repo's turn ledger the moment it happens: number, what happened, why a human was needed, provisional bucket. Feed every turn to intake as it happens, not only defects; intake's classification governs. A count held in your context is not a count (MET-05). ## Exit Green is never declared in prose. When every tier passes and the contract is met, invoke slice-exit and produce the report. If the report cannot be completed honestly, the slice is not green. ## What this skill must NOT produce - Nothing before the lock is on record; a phase started without its gate is a methodology violation the exit report must surface. - No hand-authored prototype-fidelity tests (they derive from the lock; claim-driven tests keep their test-first shape). - No invented UI for service methods the locked portion does not surface (CONF-03: record the boundary instead). - No commits without explicit direction. No kernel fixes.