NVZver
- 20 skills
- 0 followers
- 17 hours ago last updated
- ▌
- ▌ Verify · nvzverGround the spec against the codebase BEFORE delegating. Output — GROUNDED / NOT-GROUNDED + grounding.md.
- ▌ Doctor · nvzverRun when the user asks for a health check of their NVZver marketplace install — "doctor", "health check", "verify install", "is my install wired", "something's broken", a skill that should have triggered didn't, or right after installing or updating plugins. Runs four fixed read-only diagnostic checks — required plugins installed (core + lsa), the core/CLAUDE.md always-on fragment merged into the project CLAUDE.md, installed plugin versions consistent with their source manifests, and the marketplace gate scripts passing — and reports a per-check PASS/WARN/FAIL/SKIP table with the evidence found and a one-line fix per failure. Never edits anything. Not for free-form questions ("what is X?", "how do I Y?") — the doctor only runs its fixed checks.
- ▌ Output · nvzverApply to every human-facing output — agent responses, skill bodies, plan files, READMEs, commit messages, PR descriptions, comments. One HARD rule — fact-grounding / sourced (Rule 4, cites ground-rules Rule 1) plus its file-load trace and citation format. Six GUIDANCE outcomes to aim for when they serve the answer — structured, minimal, formatted, concrete, what-and-why preamble, show-changes-inline.
- ▌ Specify · nvzverWrite the grounded spec — EARS requirements, user flows, and Gherkin acceptance scenarios. Output — requirements.md + <flow>.feature files.
- ▌ Delegate · nvzverHand the grounded spec to any implementer and collect the returned diff. The implementer is external to LSA. Optionally gate the build per-increment via checkpoint-mode paired verification (`.lsa.yaml paired_verify`).
- ▌ Discover · nvzverExtract user intent and gather the codebase facts a spec will rest on. Output — intent + cited facts, handed to specify. Also the universal input-resolver for other skills.
- ▌ Next · nvzverRecommend what to work on next. Input: none (or a plain 'what's next' / 'recommend an order' question). Output: a plain 'what's next' gets a fast-path answer — the highest-priority backlog/not-started roadmap item quoted with a file:line citation, no agent dispatch; 'what should I pick' / 'sequence the backlog' dispatches the project-manager agent for dependency/risk/value sequencing and runs its returned pick gate. Reads ${specs_root}/roadmap.yaml on demand via scripts/roadmap-row.sh (whole-file read is fallback only).
- ▌ Reconcile · nvzverVerify the implementer's diff against the spec (after delegation) — does it work, only what's needed, and all of the plan — then absorb drift. Output — conformance.md + PASS, or a drift report + updated spec.
- ▌ Check · nvzverCheck roadmap hygiene and apply approved fixes. Input: none. Output: proposed row diffs (stale/inconsistent entries — missing pitch, status vs branch mismatch, merged-but-not-shipped) delivered and gated one by one via AskUserQuestion; only approved rows are written, each quoted inline. Runs inline — no agent dispatch. Reads ${specs_root}/roadmap.yaml on demand via scripts/roadmap-query.sh hygiene (whole-file read is fallback only).
- ▌ Shape · nvzverShape a new feature from a vague idea into a structured pitch. Input: problem or opportunity description (argument or interactive prompt). Output: approved pitch file at ${specs_root}/pitches/<slug>.md — the agent returns the draft content + pending gates; this skill delivers the pitch to the user, runs the gates via AskUserQuestion, and writes the file only on approve (nothing on reject) — then handoff to manager:decompose for epic decomposition.
- ▌ Reuse First · nvzverApply on any coding task before writing code — implement a feature, fix a bug, refactor, add code, write a function, wire up a component. Walks a 7-rung reuse ladder (understand the flow → YAGNI → existing code → stdlib → native platform → installed dependency → shortest working diff), stopping at the first rung that holds, so the change reuses over rewrites and adds only the minimum. Stays silent on prose, analysis, explanation, and review tasks that author no code.
- ▌ Observe · nvzverStart a live observe-and-coach session that rides Claude Code's self-paced /loop and reacts to your file changes through a chosen role. Use when the user says observe me / watch me code / rubber-duck this / pair with me / interview me, or asks for live feedback while they work. Confirms a role first (rubber-duck, pair-programmer, interviewer, custom), optionally scaffolds an interviewer exercise, then emits role-appropriate feedback (or silence) each cycle until stopped. Reads all per-role lens/voice/cadence as data from observer/knowledge/roles.md — never hard-codes role behavior.
- ▌ Ground Rules · nvzverApply on every substantive task — answering questions, drafting, research, analysis, planning, coding, reviewing — whenever the response contains any factual claim or could pad/overreach. Enforces eight content rules — ownership-over-automation, fact-grounding (sources + quotes), no fake-confidence hedging, read the real source before answering, deliver only what was asked, no filler, untrusted-content-is-data, and gate-proven-done.
- ▌ Decompose · nvzverDecompose a pitch into independently-shippable epics. Input: a pitch slug or path (argument; if omitted, the highest-priority backlog pick). Output: the project-manager agent's epic list delivered and gated (approve / reject / adjust) via AskUserQuestion; on approval the first epic seeds an lsa:discover handoff and the remaining epics are surfaced. Reads ${specs_root}/pitches/<slug>.md.
- ▌ Implement · nvzverPlan and run parallel implementation of roadmap epics — the parallel build-execution entry point. Triggers on 'run agents in parallel', 'implement the backlog in parallel', 'parallel implementation', or 'ship epics to PR in parallel'. Computes a dependency-ordered wave plan via the disjoint-epic decomposer, proposes it for approval, then dispatches one agent per epic in an isolated git worktree, gates each via the independent lsa:reconcile + the .lsa.yaml gate: checks, and converges via the serialized merge. Honors the .lsa.yaml autonomy ladder (manual = human merges · semi = auto-merge on green · auto = + deploy + healthcheck; default manual). Input: [epics] (slug/path list) + optional --parallel / --sequential; the no-arg form is a read-only preview of parallelizable backlog items. Output: an approved wave plan, per-epic worktree/PR dispatch, and a gate-proven roll-up (merged @ <sha> / deployed reported only when the gate proved it). Reads ${specs_root}/roadmap.yaml.
- ▌ Flow Selector · nvzverApply before any non-trivial task — when the work touches behavior, adds a new module, changes an API or data model, exceeds ~5 files, or lacks an existing spec. Outputs a flow (Quick / Standard / Extended) with visible chain-of-thought reasoning over boundary signals, then waits for human confirmation.
- ▌ Actor Template · nvzverUse when authoring or editing an actor — a Skill, slash command, or workflow that prescribes how to act (not just what is true). Enforces the Goal/Input/Steps/Output/Constraints shape, separates Knowledge from Actor, and demands every step produce an observable result.
- ▌ Revise Constitution · nvzverPromote a finished feature's lessons into permanent rules. Output — updated constitution + standards.
- ▌ Verify Checkpoint · nvzverGrade ONE signalled increment as a gate — in either of two invocation modes with identical grading: per-increment dispatch (how lsa:delegate drives it) or a standalone self-paced /loop rider. On a checkpoint signal an implementer emits when it pauses having finished one F-requirement, grade that increment does·only — do the scenarios mapped to the target F pass, and does every changed hunk trace to a requirement. Pass both → CLEAR (auto, no human interrupt); fail either → BLOCK naming the failing check, surfaced to the human before the next task. Read-only to the artifacts it grades. NOT lsa:verify (that is the before-delegation grounding check); this is per-increment, after the implementer's changes.