agentic

Orchestrate a feature end to end into a reviewed pull request: gather the feature and base branch, reuse an existing plan or create one, implement, commit and push, open a draft PR, run one Codex and one Claude reviewer in parallel, reconcile their findings, wait for CI, and mark the PR ready. Use when the user asks to ship a feature as a PR, run the feature PR pipeline, take a change end to end into a pull request, or orchestrate plan-implement-review for a PR. Do not use for tiny single edits, commit-only or PR-only requests, or review-only requests; dedicated skills cover those.

Create or update self-contained HTML plans and technical documents. Use when the user explicitly requests HTML, when updating an existing HTML document, or when a concrete visual or spatial layout would communicate the content materially better than Markdown. Default to Markdown for documents that fit headings, prose, lists, tables, code blocks, and straightforward diagrams, regardless of length. Do not use HTML merely to make a document polished or browser-ready; Markdown can be rendered or uploaded separately.

Upload an agent-produced plan, specification, report, or other referenced file with airplan and return a browser-ready link. Use only when the user explicitly asks for a link to a file or document the agent has produced, such as "give me a link to the plan" or "share the spec as a link". Do not use merely because a document is complete or ready for review.

Opt-in session mode that makes Codex CLI / gpt-5.6-sol the default implementer while Claude specs, reviews, and verifies. Use only when the user explicitly opts in: invokes codex-first, asks to work codex-first, says Codex should do the implementation or coding work by default, or asks to delegate the hands-on work of a session or task to Codex. Do not trigger on ordinary implementation requests that never mention Codex-led work; without this opt-in Claude does the work directly unless another explicit delegation request or workflow applies.

Ask Claude Code CLI for an independent code review of uncommitted changes, a branch diff, a commit, a PR checkout, or a specific implementation. Use when a user, agent, skill, or orchestration workflow asks a non-Claude agent to have Claude review work, when model-routing calls for a Claude review perspective, or when Claude should audit a diff for bugs, regressions, missing tests, security issues, or requirement mismatches. Do not use for small reviews the current agent can handle directly, for diffs Claude itself authored, or as a substitute for reading and validating the code yourself.

Ask Codex CLI / gpt-5.6-sol for an independent code review of uncommitted changes, a branch diff, a commit, a PR checkout, or a specific implementation. Use when a user, agent, skill, or orchestration workflow asks Claude to have Codex or gpt-5.6-sol review work, when model-routing calls for a Codex review perspective, or when Codex should audit a diff for bugs, regressions, missing tests, security issues, or requirement mismatches. Do not use for small reviews Claude can handle directly, for diffs Codex itself authored, or as a substitute for Claude reading and validating the code.

This skill should be used when the user asks to "create a pull request", "open a PR", "submit a PR", "send a PR", "make a PR for this", "PR these changes", or otherwise requests creating a pull request from current changes. Always detect, read, and use repository pull request templates before creating the PR body.

Draft pull request titles and descriptions from repository context without creating, pushing, or submitting the pull request. Use when the user asks to write, draft, rewrite, polish, summarize, or fill in PR title/body text, or when a pull request template should be completed from the current branch changes. Always detect and read repository pull request templates before drafting the PR body.

Ask Codex CLI / gpt-5.6-sol to perform read-only analysis, investigation, or data extraction over large context such as large logs, large PDFs, implementation specs, broad code searches, generated artifacts, datasets, traces, or multi-file evidence. Use when Claude needs Codex's throughput for non-destructive reading, summarization, comparison, extraction, triage, or root-cause investigation. Do not use for code edits, code review, architecture, product decisions, UX decisions, final judgement, or runtime GUI verification.

Ask Codex CLI / gpt-5.6-sol to perform local computer-use verification that needs GUI interaction, browser automation, screenshots, simulators, app launching, installation checks, runtime inspection, native app workflows, complex websites, or repeated desktop steps. Use when Claude needs observations from a running application, wants to verify UI behavior, reproduce a UI bug, collect screenshots, inspect simulator/browser/device state, or confirm a user flow. Do not use for writing code, architecture, planning, API design, code review, implementation, documentation, static analysis, or answers that can be produced from existing context.

Delegate bounded implementation work to Codex CLI / gpt-5.6-sol and then have Claude inspect, verify, and deliver the result where the work belongs. Use when the user asks Claude to have Codex or gpt-5.6-sol implement code, when model-routing calls for Codex implementation, or when a scoped task is well-defined, mechanically implementable, independently verifiable, and unlikely to require architectural, product, API, or UX judgement. Do not use for planning, architecture, exploratory debugging, ambiguous requirements, product decisions, UX decisions, or open-ended features.

Clean up stale local Git branches whose upstream remote branches are marked [gone], including removing associated Git worktrees before branch deletion. Use when the user asks to clean gone branches, remove branches deleted from the remote, prune stale local Git branches, delete branches with [gone] status, or run the clean-gone-branches Git workflow. By default, preview the gone branches with the bundled dry-run script and ask for confirmation before deleting; skip confirmation only when the user explicitly asks to delete all gone branches immediately or without confirmation.

This skill should be used when the user asks to "commit my changes", "create a commit", "commit this", "make a commit", "git commit", or otherwise requests creating a git commit from current changes.

This skill should be used when the user asks to "build a design system", "create consistent UI", "define color ratios", "set up typography system", "normalize geometry tokens", "validate visual hierarchy", "apply design constraints", or when generating frontend UI that requires systematic visual consistency. Augments the frontend-design skill with system-level visual decision rules for consistency, hierarchy, and scalable UI decision-making.

Audit, design, and incrementally improve repositories so coding agents can work reliably through legible project knowledge, runnable feedback loops, enforceable architecture rules, and cleanup processes. Use when asked to make a project suitable for harness engineering, improve agent autonomy, create an agent-ready repo structure, audit agent readiness, encode architecture or taste rules into checks, improve AGENTS.md/docs beyond instruction-file cleanup, or turn repeated agent failures into durable repo-local tooling and documentation.

Analyze GitHub pull request review feedback and turn it into a deduplicated, prioritized action plan. Use when the user asks to review PR feedback, check PR comments, summarize reviewer feedback, figure out what feedback is on a PR, determine what still needs action, dedupe repeated review comments, address PR feedback, plan PR fixes, or ask what still needs action on this PR.

Build, review, and refactor React UIs for runtime performance and responsiveness. Use when working on React apps or components that must stay fast under real load, including issues involving rerender containment, streaming state, external stores, virtualization, large lists, scroll stability, input latency, main-thread contention, persistence churn, or performance debugging.

Rebases the current branch onto the upstream default branch and checks whether newly landed upstream changes affect the branch's design. Triggers when the user says "rebase onto main", "rebase my branch", "update my branch from main", "rebase onto master", "sync with main", or otherwise requests rebasing onto the default branch.

Orchestrate a defensive vulnerability scan of a local software project by analyzing the repo at a high level, creating a bounded playbook, coordinating sub-agents to review each area, verifying findings, and managing durable vuln-scan skip decisions such as false positives or accepted risks.

Use when the user explicitly references the EDD workflow: "edd", "eval-driven development", "write evals", "feature evals", "edd-spec", "edd-impl", "edd-verify", or when working within an existing EDD feature directory (docs/features/NNN-*/). Also use when the user asks to refine or continue work on evals.md, spec.md, or tasks.md files that follow the EDD structure. Do NOT trigger for general feature requests unless the user specifically asks to use the EDD workflow.