anton-toolkit

Process and rules for a CLAUDE.md file — load this before creating, editing, auditing, or refreshing one: what belongs in it versus in .claude/rules/ or a conventions skill, the section structure, and the size limits live in references/. Covers a new CLAUDE.md, an audit of an existing one, adding a rule, and a resync after the project changed.

Creates a git commit with a Russian past-tense message from the current changes — stages files explicitly by name, one commit by default, never pushes. Use when the user asks to commit or save changes.

Systematic root-cause analysis of an error, crash, slow or wrong behavior: reproduce it, read logs and stack traces, trace the code and data, escalate to dynamic analysis, profiling, and bisection until the cause is proven — no fix is proposed on a guess. Use on an explicit request to investigate a problem, not on every mention of something not working; a sizeable fix is handed to the dev agent of the module's language, a small well-understood one is applied directly.

Interview-driven feature planning: turns a vague idea, ticket, or brief into a confirmed, actionable spec saved to docs/plans/ before any code is written, then offers handoff to a dev agent. Use when the user wants to plan, refine, or clarify a feature before implementation; skip it when the spec is already confirmed or the change fits in one sentence (delegate straight to java-dev, kotlin-dev, python-dev, go-dev, or frontend-dev, or just implement it), use debug for a bug report, and improve-plugin / create-plugin / extend-plugin for a plugin skill or agent. Runs in the main conversation; the interview is not delegated to agents.

Designs parts and assemblies in FreeCAD through the freecad MCP server: the user says what to make, the skill models it parametrically, the user opens FreeCAD and inspects it himself, then asks for changes. Covers modelling for 3D printing, milling, turning, welding and sheet furniture, assemblies with working joints and driven motion, load analysis, drawings and STL/STEP export. Use for any request to model, design, or correct something in FreeCAD or CAD. Not for CAM toolpaths or G-code. Runs in the main conversation — the dialog is not delegated to an agent.

Frontend (React/TypeScript) conventions of this toolkit: no business logic on the client, one hook per page, LLM-friendly naming, types, components and comments, non-generic visual choices, and the build/lint/test done criteria. Loaded automatically when working with frontend files; also preloaded into the frontend-dev agent.

Go conventions of this toolkit: error handling, context and goroutine discipline, interfaces and exports, the zero-comment policy, module and dependency rules, and the gofmt/build/vet/lint/test done criteria. Loaded automatically when working with Go files; also preloaded into the go-dev agent.

Java conventions of this toolkit: Spring Boot project patterns, mandatory braces, one guard per loop, MapStruct mapping, decomposition, and the compile/test done criteria. Loaded automatically when working with Java files; also preloaded into the java-dev agent.

Five concrete coding principles (think before coding, simplicity first, surgical changes, goal-driven execution, reuse before writing) with correct/incorrect pairs; applied by every dev agent and whenever code is written.

Kotlin conventions of this toolkit: Spring Boot module patterns, the no-comment and brace rules, idiomatic-Kotlin rewrites, null safety, atomic writes, detekt discipline, and the compile/detekt/test done criteria. Loaded automatically when working with Kotlin files; also preloaded into the kotlin-dev agent.

Reviews all changes on the current branch against the project's own documentation and returns ready-to-paste MR comments with file:line locations. Run with /mr-review; for a self-contained best-practice review use code-reviewer instead.

Python conventions of this toolkit: PEP 8 with full type hints, no comments or docstrings by default, error and async rules, package-manager discipline, and the ruff/mypy/pytest done criteria. Loaded automatically when working with Python files; also preloaded into the python-dev agent.

Rust conventions of this toolkit: errors as Result, the unsafe and panic rules, ownership instead of shared-mutability workarounds, the zero-comment policy, flat data layout in hot code, crate and dependency rules, WebAssembly specifics, and the fmt/clippy/test done criteria. Loaded automatically when working with Rust files; also preloaded into the rust-dev agent.

Takes one work task and drives it to reviewed, tested and QA'd code in a single run: plans it through feature-planner, routes every affected module to the dev agent of its stack, then runs the build pipeline — code with tests, one review of the whole diff with adversarial checking of the blockers, fix rounds, local stand, end-to-end QA, final review mark — and hands back what only the user can decide. Use when the user gives a task, a ticket or a plan to implement. For planning without building use feature-planner, for an unexplained failure debug, for a whole system's documentation system-designer, and for the Logos project logos-build. Runs in the main conversation; the planning dialog is not delegated to agents.