skills

Execute exactly one Linear issue end-to-end with branch safety, repo-discovered gates, self-QA, adversarial review, decision logging, single-source report and PR-body rendering, Ready PR creation, and autonomous post-PR review-comment handling. Use when the user says /ship, ship issue, ship Linear issue work, or asks to complete a specific Linear issue such as ENG-123; reject GitHub issues, plan phases, and vague next-work requests.

Per-feature Neon database branch for each git branch in a Drizzle + Vercel project. Schema changes use `db:push` (no migration files on feature branches); migrations are auto-generated and validated post-merge against a throwaway prod clone, then applied to prod with deploy gated on success. Use when setting up or operating a project where multiple agents/worktrees need isolated DB schemas without conflicting migration files. Triggers on "isolated db branches", "neon branching", "per-branch database", "db isolation", "drizzle migration conflict", "neon worktree", "/isolated-db-branches".

Execute one implementation slice end-to-end from either (1) a PRD plus plan plus phase, where PRD and plan may be files or GitHub issues, (2) a GitHub PRD issue plus optional work issue, or (3) a Linear issue with its project/docs as PRD context. Use when the user says "/work", "do work", "do issue work", "do linear work", asks to implement a planned phase, asks to pick/execute the next GitHub slice from a PRD, or asks to complete a Linear issue. Always includes dependency intake, clarifications, feature branch, implementation, mandatory gates, adversarial review, Playwright self-QA or fallback evidence, local conventional commit, 2-page HTML report, opening the report, pushing the branch, and opening a Ready PR whose body mirrors the report.

Run a multi-round adversarial code review with two independent AI reviewers that cross-examine each other's findings. Iterates until only nitpicks remain (max 3 rounds). Use when you want deep, debate-style review that catches issues a single pass misses. Triggers on "adversarial review", "debate review", "cross-review", "deep review".

Review and resolve unresolved GitHub PR review threads. Use when the user wants to process PR comments, address requested changes, triage reviewer feedback, reply to invalid feedback, push fixes to the PR branch, and resolve review threads safely. Adapts quality gates to the repository instead of assuming a fixed package manager or test suite.

Analyze text readability with Flesch-Kincaid, Gunning Fog, SMOG, and other metrics. Returns objective scores with interpretation and recommendations.

Write a technical retrospective grounded in artifacts the user can verify — commits, PRs, ADRs, tests, logs. Use only when the user has a specific past incident with concrete evidence to cite. Refuses to fabricate stages when evidence is missing. Triggers on "engineering retrospective", "lessons-learned article", "technical postmortem write-up", or "blog post about a bug, incident, or fix".

Log accomplishments, solved problems, learnings, and milestones to Obsidian daily notes via the Obsidian CLI. Also answers questions about the Obsidian CLI. Triggers on "/log", "log accomplishment", "log what I did", "work log", "obsidian cli", "daily note", or when user wants to record something noteworthy they did.

Software engineering principles for code review, planning validation, and learning. Covers SOLID, DRY, YAGNI, design heuristics, testing, API design, architecture patterns, and React best practices. Use when reviewing code, validating plans/specs, injecting review criteria into council sessions, or when user asks about a specific principle. Triggers on "principles", "code review checklist", "SOLID", "DRY", "design heuristics", "review criteria", "best practices", or any named principle like "Law of Demeter", "Liskov", "YAGNI", etc.

Dispatch work to an external CLI agent (Claude Code, Codex CLI, Gemini CLI, or GitHub Copilot CLI) by writing a prompt and launching the agent via background Bash tasks. Use when orchestrating multi-agent workflows, sending work for external review, parallel task fan-out, or the "review council" pattern where multiple agents review a plan in parallel. Triggers on words like "dispatch", "fan-out", "review council", "send to agent", "launch agent", or when orchestrating parallel CLI agent work.

Send a document (plan, spec, PRD, ADR, code) to a review council of 2+ AI agents (Claude Code + Codex CLI by default, optionally Gemini CLI and Copilot CLI) for independent parallel review. Synthesizes feedback, auto-revises the document, and re-reviews if council members had high variance. Use when wanting multi-agent review, "review council", peer review from multiple AIs, or validating plans/specs/code before finalizing.