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).
Installs into .claude/skills of the current project.
Are you the author of Implement?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/teklabsdigital-implement)
---
name: implement
description: 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).
---
# 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.