Software Development Skills

Run a continuous audit-fix-audit loop that scores a codebase and automatically fixes gaps until a target score is reached. Use when the user asks to "get us to 10/10", "keep improving until perfect", "run a continuous improvement loop", "audit and fix automatically", or wants an autonomous cycle of scoring, generating fix tasks, executing them, and re-scoring.

The native task command centre for a GRU953-Studio project — plan, track and control work with pause, resume, stop, skip-now and schedule-for-later. Defines the task state machine over PROGRESS.md (adding paused/skipped/scheduled to todo/doing/done/blocked), the live plain-English STATUS-BOARD.md, how each /studio control command changes state, how "schedule for later" uses the session's scheduler (degrading gracefully when none exists), and the self-contained HTML dashboard. Use whenever the user wants to see, pause, resume, stop, skip or schedule work.

The plain-text memory system that makes every GRU953-Studio project resumable — in the same session or a brand-new one, days later. Defines the exact files, the ▶ RESUME HERE pointer, the secrets-scan-before-write rule, the resume rehearsal, and the rule that memory is local-only and never ships. Use at the start and end of every session, after every stage or task, and whenever the studio needs to read or write project memory. The memory-keeper role performs the reads and writes; this skill is the protocol it follows.

The protocol for recommending an existing Claude Code skill or plugin when a task would clearly benefit from one GRU953-Studio doesn't build natively (a language server, a specialised third-party integration, a niche workflow tool) — never bundled into GRU953-Studio itself, always confirmed with the user before anything installs. Use whenever researcher is woken for this reason, and whenever project-lead needs to present or act on its recommendation.

Use before a user's very first GRU953-Studio project — the one-off "getting to know you" setup that runs once, never on later projects. Confirmed 2026-07-10 as a separate step rather than folding it into the first project's own interview.

The anti-drift spine for long, multi-session, complex builds. Defines Dev-Memory/FOCUS.md (the one-glance anchor that survives a summarised or brand-new session), the re-orientation ritual run at every session start and every stage boundary, the drift check that stops work not traceable to a confirmed requirement, and Dev-Memory/REQUIREMENTS.md (the traceability matrix, mechanically audited by hooks/traceability-check.mjs). Load and follow as a standing rule on every project, alongside dev-memory. Use at the start of every session, at every stage boundary, and before starting any task.

The C++ ecosystem pack — the exact build, test, lint, format and dependency commands, plus the idioms, that the cpp-developer agent uses. Load when a task is implemented in C++. Covers CMake/CTest, clang-tidy, clang-format, and vcpkg/Conan dependency and licence norms (best-effort).

The C / .NET ecosystem pack — the exact build, test, lint, format and dependency commands, plus the idioms, that the csharp-developer agent uses. Load when a task is implemented in C. Covers the dotnet CLI, xUnit/NUnit, analyzers, dotnet format, and NuGet dependency/licence norms for Windows and cross-platform .NET.

The Dart & Flutter ecosystem pack — the exact build, test, analyse, format and dependency commands, plus the idioms, that the flutter-dart-developer agent uses. Load when a task is implemented in Dart or Flutter. Covers pub, flutter/dart analyze, dart format, and pubspec dependency/licence norms.

Go 1.23+ development specialist covering Chi, GORM, and concurrent programming patterns. Use when building high-performance microservices, CLI tools, or cloud-native applications.

The Java ecosystem pack — the exact build, test, lint, format and dependency commands, plus the idioms, that the java-developer agent uses. Load when a task is implemented in Java. Covers Maven and Gradle, Checkstyle/SpotBugs, google-java-format/Spotless, and dependency and licence norms.

The Kotlin ecosystem pack — the exact build, test, lint, format and dependency commands, plus the idioms, that the kotlin-developer agent uses. Load when a task is implemented in Kotlin. Covers Gradle, ktlint/detekt, and Gradle/Maven dependency and licence norms.

The Python ecosystem pack — the exact build, test, lint, format and dependency commands, plus the idioms, that the python-developer agent uses. Load when a task is implemented in Python. Covers pytest, ruff, black/ruff-format, mypy, and pip/requirements/pyproject dependency and licence norms.

Rust 1.92+ development specialist covering Axum, Tokio, SQLx, and memory-safe systems programming.

The Swift ecosystem pack — the exact build, test, lint, format and dependency commands, plus the idioms, that the swift-developer agent uses. Load when a task is implemented in Swift. Covers SwiftPM/xcodebuild, XCTest, SwiftLint, swift-format, and Swift Package Manager dependency/licence norms for iOS and macOS.

The TypeScript ecosystem pack — the exact build, test, lint, format and dependency commands, plus the idioms, that the typescript-developer agent uses. Load when a task is implemented in TypeScript. Covers tsc, jest/vitest, eslint, prettier, and npm dependency/licence norms for web and cross-platform (React Native/Electron/Node) work.

The token-cheap recall layer over Dev-Memory — a compact structured INDEX.md (the cheap first read) plus a plain-text knowledge graph GRAPH.md of linked entities (tasks, requirements, decisions, files, lessons) with typed links, so a session recalls only what the current task needs instead of re-reading everything. Zero external dependencies; optional local semantic re-rank via Ollama only when it is already present. Load and follow as a standing rule alongside dev-memory. Use at session start (recall) and after any task/decision (record). Audited by hooks/memory-integrity.mjs.

Turns a confirmed design into an ordered list of small, independently-verifiable "micro-tasks" — each with one acceptance criterion and the exact command that proves it, scaled by project Tier — so nothing moves from design to code without a clear, checkable plan. Use whenever architect finishes a design, and whenever project-lead/builder/tester need a task's exact acceptance criterion, verification command, or execution order.

The protocol for using Ollama (a free tool for running AI models locally, no cloud needed) two ways — as an alternative AI backend ai-developer can offer for an app GRU953-Studio builds, and as a standing capability any role can reach for as a free, private, independent local second opinion. Always confirms with the user before installing Ollama itself or pulling any model — every time, no exceptions. Use whenever ai-developer designs an AI feature, and whenever a role wants a local-model pass instead of (or alongside) its own review.

Turns the confirmed design into an MVP-first, then progressively-enhanced build roadmap — Phase 1 is the smallest core that actually works, Phases 2..N add the rest in order. PLAN.md/PROGRESS.md gain a Phase column; each phase is independently shippable and ends in a backup checkpoint. Use at the Prototype/Plan boundary and whenever the build plan is organised or revised.

The GRU953-Studio protocol for publishing a finished MVP privately to the user's OWN GitHub account, with an explicit separate step to make it public later. Referenced by name from publisher.md and the /studio-publish command, which read this file directly — never auto-invoked by Claude on its own initiative (see disable-model-invocation below), since publishing is a side-effectful, GitHub-pushing action only the user's own explicit trigger should start. Covers pre-flight secret/vulnerability/licence scanning, the signed-in user's own author identity, attribution cleanup, private repository creation and the final report.

The Prototype stage — before any real code, build a self-contained clickable HTML "warframe" (wireframe prototype) of the app plus the phased build plan, then get explicit user approval of both at a hard, blocking gate. Only after approval does Plan/Build begin. Use right after Design, on every project with a user interface; for a pure CLI/library, a lightweight text/ASCII walkthrough stands in for the visual warframe.

The GRU953-Studio lean-coding rule set (YAGNI — "You Aren't Gonna Need It"). Use whenever the studio writes, reviews or trims code, whenever a builder is about to add a file, function, dependency or abstraction, and whenever the user says "keep it simple", "minimal", "lean", "no bloat", or "just the basics". Every line of studio-produced code must pass this ladder.

每完成一个 plan 步骤自动 git commit,记录到 .ddt/checkpoints.log;支持 git revert 回滚到任意 checkpoint。auto-loaded by build-api / build-web / fix 等 IMPLEMENT phase。DDT 步骤级 git commit…