Coding & Development

Implement Astro interaction patterns such as navigation, dialogs, tabs, filters, cards and theme switching. Reuse framework primitives while preserving each pattern’s behaviour contract.

Write and explain SQL queries for a Postgres database.

Review a pull request for correctness, safety, speed and readability.

Accessible forms with react-hook-form, zod and shadcn/ui Form, covering inline errors, pending and disabled states, success feedback and server-side validation with shared schemas.

Review software changes, diagnose bugs with evidence, and assess regressions, lifecycle failures, permissions, feature flags, performance, and test coverage. Use for requested code reviews, difficult bug investigations, and the final review of substantial implementation work. Scale to the changed behavior; do not turn routine edits into whole-project audits or edit during a review-only request.

Build and verify lightweight, visually ambitious 3D Web games using Three.js/WebGL2, procedural geometry and materials, synthesized audio, and compact gameplay systems. Use when the user requests a small single-file browser game, an offline HTML 3D game, AAA-style Web graphics, procedural game assets, or size/performance optimization of such a game. Includes rendered visual validation, gameplay checks, and honest hardware performance gates. Not a promise of commercial AAA content scale.

Testing profesional de aplicaciones React: Vitest para unidades y hooks, React Testing Library para componentes con criterio de queries por rol, MSW para mockear red a nivel de API, Playwright para end-to-end, cobertura de interacciones reales, cobertura de componentes y los limites de cada nivel. CARGAR cuando hay que escribir tests de una interfaz, decidir que mockear, configurar Vitest o Testing Library, montar un componente o escribir un flujo end-to-end.

Finds and fixes game frame-rate problems by measurement in Godot, Unity, Unreal or Roblox: CPU- vs GPU-bound, then the biggest fix (draw calls, pooling, GC spikes). Use for low FPS, stutter, a first-use hitch, or requests to optimize or profile. Not for build size (`unity-build-size`), physics jitter (`physics-tuning`) or network lag (`roblox-networking`, `godot-multiplayer`).

Designing a library — a published package, or an in-app module from the moment it has a second consumer. The implementation stays editable forever, while the public API, the dependency set, the performance characteristics and every observable behaviour freeze at first external use (Hyrum's law). Five duties to the caller: take no control, add no hidden cost (zero-cost abstraction as the bar), hold no hidden state, hide no failure, block no escape. Use when designing or reviewing a public API, when extracting shared code into a module, when asked to "make this reusable", when a signature-compatible change alters an observable behaviour, and when an application is being shaped around a framework's types and lifecycle.

Deriving test cases, triggered by the moments that produce weak ones: about to write a test list that is the happy path restated, about to derive tests by reading the implementation they check, about to test the bound the specification names and not the other one, about to enumerate the roles the specification lists and skip the caller it forgot, about to vary one condition at a time from the happy path, about to draw a state or decision table without tying each row to a test, about to write an expected result with "or" in it, about to skip the error paths as unlikely, about to leave a path unimplemented and untested, or about to grade a suite as adequate by reading it. A test derived from the implementation passes by construction; coverage counts classes, not tests. Use when writing a test plan or list, when a use case is the specification in hand, when adding tests to existing code, when reviewing a suite for gaps, and when a suite is all green and proves little.

Find every place a repo calls an LLM and write a route map (routes.json). Use when the user says "audit my LLM usage", "why is my LLM bill high", "make my LLM calls faster", "where do we call the model", "map our LLM calls", or starts any LLM cost or latency review. Run this first; llm-capture and llm-diagnose need its output.

Use for "interrogate", "adversarial review", "multi-model review", "challenge this", "second opinion", "second look", "2nd opinion", "what could this break", "find blind spots", or "tear this apart". Multiple LLM reviewers challenge changes from independent angles; only findings a run proves survive.

Apply clean-code principles when writing, reviewing, or refactoring code to improve readability, maintainability, testability, and extensibility. Use for code-smell reviews, naming and function design, duplication removal, class responsibilities, and behavior-preserving cleanup across projects and languages.

Revisar diffs, commits, branches, Pull Requests e alterações locais com foco em defeitos, regressões, segurança, desempenho, testes e manutenibilidade. Usar quando Codex precisar realizar revisão ou auditoria de código, avaliar a qualidade de uma mudança, fornecer feedback técnico ou definir critérios de aprovação, especialmente em projetos PHP/CakePHP e aplicações web.

Write Appium Java mobile automation tests (Android and iOS) from manual test cases, or migrate existing Python/Appium (pytest) tests to Java - using ThreadLocal drivers for parallel devices, explicit condition-based waits, cross-platform Page Object Model, W3C gestures, TestNG and Allure. Use this whenever the user pastes mobile/app test cases and wants them automated, shares Python Appium code and wants it in Java, asks to "convert pytest to Java", "port to java-client", "automate this Android/iOS flow", "write Appium tests", or wants to scaffold an Appium Java framework, even if they don't say "Appium" but are clearly automating a native or hybrid mobile app in Java.

Chooses Python callable ownership, including lambda versus named function and method versus module placement, plus properties, closures, cache lifetimes, construction/subclass hooks, and deterministic cleanup. Use when those design decisions affect reuse, binding, tracebacks, state, or lifetime. Do not use for ordinary calls to partial or other higher-order helpers, browser/distributed/external caches, dependency maintenance, or isolated syntax.

Convert manual test cases into Playwright Java automation tests using a production-style framework - ThreadLocal browser sessions for parallel runs, condition-based waits (no sleeps), Page Object Model, TestNG and Allure reporting. Use this whenever the user pastes or uploads manual test cases (table, Excel rows, Jira/TestRail export, Gherkin, or plain numbered steps) and wants them automated in Java, asks to "automate these test cases", "write Playwright tests for this", "convert to POM", "add page objects", or wants to scaffold a Playwright Java framework, even if they don't say "Playwright" but are clearly working in a Java web-automation project.

Builds and reviews bounded-memory Python file-processing CLIs with explicit path bases, argparse contracts, and iterable- or callback-driven tqdm progress. Use when file scale, path ownership, progress semantics, or CLI behavior needs design. Never activate for tiny fixed fixtures that can be read wholly, swapping an already-chosen progress library, translation, documentation-only edits, dependency updates, web APIs, GUIs, shell-only scripts, or incidental Path/tqdm/argparse terms.

Carry a software task end to end, from an isolated worktree through implementation, verification and delivery to a resumable handoff. Use for substantial feature, bug-fix or refactor work, especially when the user hands off a task and expects a finished, verified result without supervising it.

Designs Python type annotations, domain objects, validation boundaries, dataclass/Pydantic models, invariants, interfaces, fixed choices, and object-copy semantics. Use when choosing ABC for owned nominal implementations versus Protocol for third-party structural compatibility, or when changing model behavior and boundary contracts. Never activate for dependency/version updates, translation, documentation-only edits, counting references, isolated syntax, or incidental modeling terms.

Review kualitas seluruh repo: bug, celah keamanan, titik rusak di bawah beban nyata, kode berisiko tanpa test, jalur lambat, dan apa yang perlu dihapus, digabung, atau dipecah. Laporan satu kali, berurutan, tanpa mengubah kode. Gunakan untuk "review codebase", "review repo", "cari bloat", "apa yang bisa dihapus", /ponytail-review.

Mode senior dev malas: perubahan terkecil yang menyelesaikan tugas sepenuhnya, dengan balasan yang dipahami sekali baca. Gunakan pada tugas coding (tulis, fix, refactor, review, pilih dependency) atau saat pengguna bilang "ponytail", "be lazy", "simplest solution", "yagni", atau mengeluh soal over-engineering. Level: lite, full (default), ultra.

Apply a lightweight intent, complexity and resource assessment to every task. Handle simple work directly; route substantial work to available tools or agents with dependency-aware ownership, persistent checkpoints, integration and verification. Recover interrupted runs from actual repository evidence.

Create, review, or revise auditable software estimates for AI-assisted delivery, agentic SDLC workflows, agentic products, or hybrid agentic programs from documents, spreadsheets, repositories, tickets, screenshots, links, or pasted requirements. Use when scope must be analyzed against the actual technology and architecture, related requirements must be intelligently merged or separated, material clarifications must be asked before estimating, client Solution, technical Dev Solution, AI Use, and Assumptions text must be written into workbook task rows, or manual, AI-assisted, blended, three-point, cost, duration, calibration, and formula-audit outputs are needed across NetSuite, SuiteCommerce, .NET, web, mobile, data, cloud, and mixed stacks.