Claude Skills by vraj-ai

Tear down a converged item or whole deliverable as a read-only different-model judge, returning falsifiable findings and a blocking verdict. Use for optional hardened T0 review, optional T2 whole-deliverable review, and the mandatory final T3 code-level gate.

Run independent multi-model research, evidence-based debate, voting, and scoped T0/T1 reviews without rubber-stamping. Use when goals encounters a genuinely contested research item, Phase B needs backlog sanity checks, an item needs independent T0 reviewers, or a milestone needs one rotating integration reviewer.

Drive a plan document to verified completion through a resumable single-writer backlog, named Worker Role research and review, worktree-isolated parallel builds, milestone gates, and a final adversarial teardown. Use when the user invokes /goal, asks to execute an architecture plan autonomously, resumes a goal, or needs a large issue set delivered milestone by milestone.

Drive a published spec to verified, pushed completion through a resumable backlog, worktree-isolated parallel builds under a lazy-senior-dev ladder, gate-first review, a milestone reviewer, and a final adversarial teardown. Use when the user invokes /ship, asks to build out a spec autonomously, or resumes a ship run.

Legacy coding stage from the former planner-coder-debugger-reviewer fleet, preserved for explicit opt-in use. Use only when the user invokes /legacy-coder or deliberately requests the old GitHub Projects stage workflow instead of goals.

Legacy debugging stage from the former planner-coder-debugger-reviewer fleet, preserved for explicit opt-in use. Use only when the user invokes /legacy-debugger or deliberately requests the old GitHub Projects stage workflow instead of goals.

Legacy planning stage from the former planner-coder-debugger-reviewer fleet, preserved for explicit opt-in use. Use only when the user invokes /legacy-planner or deliberately requests the old GitHub Projects stage workflow instead of goals.

Legacy review stage from the former planner-coder-debugger-reviewer fleet, preserved for explicit opt-in use. Use only when the user invokes /legacy-reviewer or deliberately requests the old GitHub Projects stage workflow instead of goals.

Pricing, limits, and margin gates for a product whose cost of goods is model inference — compute cost per action, cost per active user, and the breakeven limit before choosing a price, then encode the resulting caps as testable invariants. Use when pricing an AI feature or subscription tier, setting rate/usage limits, evaluating whether a plan loses money on heavy users, or estimating the inference cost of a proposed feature.

Full-stack audit methodology for a whole codebase or app rather than a single diff — establish the boundary and baseline, sweep layer by layer (entry points, domain, data, integrations, auth, tests, ops), rank findings by blast radius, and emit each finding as a ready-to-implement ticket with a Verification-command. Use when asked to audit an app or codebase, assess a project's health, review before a launch, or produce a prioritized defect backlog.

Deliver a ticket under explicit external constraints — token/spend budget, a live data migration, restricted git authority, or limited tracker access — by naming every constraint up front and choosing a delivery path that respects them. Use when budget is capped, when a change touches production data or costs money to run, when push/PR authority is restricted, or when tracker writes are unavailable and work must hand off through a file instead.

Stage protocol for the preserved legacy GitHub Projects workflow, where Project Status is authoritative and work crosses separate planner, coder, debugger, and reviewer sessions. Use only with legacy-planner, legacy-coder, legacy-debugger, or legacy-reviewer.

Audit whether a project's stated invariants are actually ENFORCED and MEASURED in the shipped path, rather than merely mentioned in a plan or asserted in a comment — for each invariant find the enforcement point, the test that proves it, and the signal that would catch its violation in production. Use when reviewing a diff against locked invariants, as net 3 of debugger's audit, before accepting a security/performance/privacy claim, or when the user asks whether a constraint is real.

Fan out work to several subagents at once on lanes you drew, with the parent holding every gate and performing all commits and tracker writes. Use when the user has explicitly authorized parallel or batch work, or when another skill needs the lane-and-gatekeeper protocol. Also routes to the right fan-out skill for plan tasks, dependency-ordered ticket ranges, and shared-tree mechanics.

Run the preserved legacy planner, coder, debugger, and reviewer effort across separate stage-parent sessions with an evidence gate between every stage. Use only when deliberately coordinating the legacy four-stage workflow rather than the active goals pipeline.

Test-first method for integrations with external providers — queues, signed webhooks, idempotent lifecycle operations, and owned artifacts — where the failure modes are delivery-order, replay, partial failure, and forged payloads rather than plain logic bugs. Use when implementing or auditing a webhook receiver, a job queue consumer, a payment/billing lifecycle, an upload/artifact pipeline, or any code path where a third party can call you back.

Safely fan out subagent tasks into one shared git checkout — assign explicit file lanes, forbid subagent commits/pushes by default, and have the parent re-run every gate because subagent summaries are claims not evidence. Use when delegating implementation work to subagents that share a worktree, when writing a subagent brief, or when a parent agent must review returned work. Prefer separate worktrees when the work genuinely conflicts.

Invariants for working in a git checkout that another agent or human may be writing to at the same time — capture a baseline, preserve dirty paths you do not own, treat mid-run file changes as a concurrency signal, stage narrowly, and re-check ownership before every external write. Use whenever parallel agents, background tasks, or a human share one worktree, before any commit/push/tracker write in a shared tree, or when files change underneath you mid-run.

Create and verify role agents for the preserved legacy planner-coder-debugger-reviewer workflow. Use only when maintaining the legacy four-stage profile rather than the active goals/council/parallel OpenCode profile.

Unstick a pipeline that is thrashing or reporting false progress — detect the false-green smells, classify the blockage as implementation defect vs missing structural prerequisite, and either repair with coder discipline or write a truthful blocker and spawn the prerequisite ticket. Use when a ticket has been retried repeatedly with no new evidence, when a worker reports success but nothing changed, when a ticket sits in Coding but cannot actually land, or when the tracker state and the repo ...

Deliver an explicitly authorized RANGE of tickets in sequential waves, where each wave respects blocker ordering, the parent verifies every wave before starting the next, and a red wave halts the batch. Use when the user authorizes a GitHub issue range such as "do #201 through #208", when a plan's children must land in dependency order, or when a batch is too large or too coupled for a single parallel fan-out.

Strict one-ticket TDD for pipeline work — lock the Verification-command before editing, drive red-green at the highest meaningful seam, keep provider calls behind test-safe seams that the production path actually uses, and finish with native checks (typecheck, lint, build) plus the gate re-run. Use when implementing a Coding ticket, especially one touching external providers, persistence, or auth, and whenever coder needs the detailed implementation loop.

Control a connected browser when the user explicitly asks for navigation, visible-state inspection, interaction, screenshots, or local web testing; do not auto-invoke for URL context or semantic connector work.

Run or generate a gauntlet loop for one-shot prompts, UI work, writing, research, and implementation. Use when the user says gauntlet, beat this reference, polish until better, make a one-shot prompt, compare against a real product, or loop with a harsh critic.

Run a deliberate GitHub repository, issue, pull-request, review, CI, or publish workflow when the user explicitly invokes this skill.

User-invoked interview that stress-tests a plan and writes glossary, architecture, and ADRs into CONTEXT/. Use when the user runs /grill.

User-invoked, at-your-own-risk mode. Finishes any task; ignores other skills' judgement stops; leaves commit, push, and merge to the human; waits for a push then a PR merge.

Interactive Herdr setup and orchestration for a Grok pane coordinating Codex and three non-OpenCode Council agents.

User-invoked. Sets up a project's tracker and CONTEXT/ once, then turns a grill into a spec and tracer-bullet tickets. Use when the user runs /issues.

Run a coding task as a closed maker→checker loop with an explicit done-condition and rubric, iterating until the goal is verifiably met instead of stopping after one pass. Use when the user wants an agent loop, "loop engineering", a self-verifying build, maker-checker, "keep going until it's done/passes", autonomous iteration on a feature or bug, or asks to drive a task to a measurable finish line.

Run one task through several model/provider agents in isolated worktrees, relay cross-critiques, and compare the final artifacts. Use when the user invokes /multi-agent-review, wants competing implementations/plans/fixes, or explicitly runs the preserved legacy fleet's multi-model mechanism.

Apply an explicitly requested, reversible maintenance pass to a Pi installation, skills, extensions, themes, and settings.

Run an explicitly requested maintenance pass over Pi usage data, indexes, reports, and retention checks.

Audit a whole repository for over-engineering; use when the user asks to audit a codebase, find bloat, or invokes ponytail-audit. One-shot and read-only.

Harvest every ponytail shortcut comment into a debt ledger; use for ponytail debt, deferred shortcuts, or a one-shot ledger report.

Show Ponytail's measured impact as a one-shot scoreboard; use for ponytail gain, savings, impact, or scoreboard requests without changing files or mode.

Display a one-shot quick-reference card for Ponytail modes, skills, commands, and updates; use for ponytail help or command questions.

Review a diff exclusively for over-engineering; use when the user asks what can be deleted, whether code is over-engineered, or invokes ponytail-review.

Forces the laziest solution that actually works; use on coding tasks or when the user says ponytail, be lazy, simplest solution, YAGNI, do less, or complains about over-engineering, bloat, boilerplate, or unnecessary dependencies. Do not use for non-coding requests.

GitHub Actions PR review bot with Greptile parity — confidence 0-5, risk gates, checks annotations, and fix loop. Install via setup-vskills.

Commit and push a verified handoff plus its artifacts, only under explicit authority, and prove the push happened by reading the remote SHA back. Use as the final delivery step after artifacts are ready, when the user asks to push work, or when a handoff needs to reach the remote. Refuses to claim success without remote proof and never force-pushes or commits secrets.

Turn a repo's docs folder into a retrieval graph — router, generated indexes, state file — and open it as an Obsidian vault.

Sets up this skills repo on a new machine — installs the skills with the vskills CLI, then regenerates the local-only context docs (CONTEXT.md, docs/) that are deliberately not published in the public repo.

User-invoked session close. Syncs CONTEXT/ and the issue tracker, writes a handoff, then commits and pushes under /snapshot authority. Use when the user runs /snapshot.