codex-claude-code-starter-pack

Run a decision-grade, multi-source research pass that verifies claims across official docs, implementation evidence, field evidence, and counter-evidence, then turns them into a source-backed memo or an adopt/experiment/watch/reject decision. Use when the task needs multiple source layers, counter-evidence, contradiction handling, or a source-of-truth comparison across tools, vendors, workflows, repos, standards, or market/technical claims. Do not use for quick fact lookup, a single-doc answer, routine digest review, simple feature comparison, code review, or implementation-only work.

Convert scanned Hebrew or mixed Hebrew-English PDFs into source-faithful Markdown with RTL-safe structure, page-level provenance, uncertainty tracking, and OCR QA. Use when Codex needs to read Hebrew PDFs, especially scanned, rotated, or medical documents, and preserve wording, dates, tables, dosages, and mixed-direction text instead of producing only a summary.

Elevate the visual quality of websites, apps, dashboards, presentations, and other visual outputs after the first functional pass. Use when the output works but still feels generic, flat, confusing, or unfinished.

Create distinctive, production-grade frontend interfaces with high design quality. Use this skill when the user asks to build web components, pages, or applications. Generates creative, polished code that avoids generic AI aesthetics.

Use when the user wants a presentation, slide deck, lesson slides, business deck, PowerPoint, or a short visual talk. Create a clear outline first, then build a simple deck that is easy to read and present.

Use when the user wants to open a real site or app in a browser, click through it, take screenshots, verify a flow, or debug what the UI actually does.

Use when the user wants to preview or publish a website or app and get a live link. Prefer preview deploys unless the user explicitly asks for production.

Review a website or app UI against modern web interface guidelines. Use when the agent is building or auditing layout, accessibility, forms, interaction states, or overall UX quality.

Use before writing code or taking action when the request is still fuzzy. Compare a few real options, recommend one, and keep the answer simple enough for a beginner to follow.

Use after a plan exists. Execute one step at a time, verify each step, and replan when the evidence changes.

Use at the start of substantial Codex work to choose the right model style: strong reasoning for planning, implementation-focused for coding, and faster loops for small iterative edits.

Use when building or reviewing a React or Next.js UI. Keep components small, state simple, and user-facing flows understandable.

Use when a bug, failure, or confusing behavior appears. Focus on root cause, not random edits.

Use when building features with behavior that can be checked. Work in a small red-green-refactor loop instead of writing a large untested block.

Use when parallel work or risky changes need isolation. Create a separate worktree so one task does not break another.

Use before saying a task is done. Run the real check, read the output, and only then claim success.

Use when the work is large enough to need sequencing. Break it into small verifiable steps with clear done criteria, risks, and checks.