skills

Serious engineering work with a reviewed plan and adversarial review gates; the required first argument word picks the mode. `build` — features, root-caused fixes, cross-cutting changes (trivial ones skip planning, never the diff gate). `refactor` — consolidate one accreted component, behavior-preserving. `sweep` — unattended overnight cleanup of a named area, shipped as draft PRs. `experiment` — metric-driven experimentation: approved hypothesis tree, autonomous batch of runs, next-batch report. `followup` — work recorded out-of-scope findings with the user's decisions; no required argument. Invoke explicitly, e.g. `/tokenmaxxer refactor <component>`.

Guidelines for ML projects. Use when training or evaluating models, building ML pipelines, running experiments, or working with datasets/parquet files and CatBoost/PyTorch/Polars code.

Build full-stack web applications using Rust (Axum + SQLx) with server-rendered frontend patterns using HTMX + Alpine.js or DataStar, plus Neon (serverless PostgreSQL). Use when asked to create web apps, CRUD apps, dashboards, forms, or any stateful web application. Triggers on requests like "build a todo app", "create a voting app", "make a dashboard", "build a blog", etc.

Prove and counter-prove a claim with parallel agents before concluding. Use for architecture claims, bug hypotheses, performance claims, refactor safety, review judgments, and "is this actually true?" questions.

Evidence-first debugging and root cause investigation. Use for triaging bugs, test failures, incidents, performance regressions, flaky behavior, integration failures, or unexplained behavior before proposing fixes.

Design, build, audit, and improve the harness around an LLM agent — its tools, loop, permissions, context, skills, and evals. Use when building, reviewing, or debugging any agentic system: an agent runtime or loop, agent-facing tools, an MCP server, an approval/permission flow, a context or memory strategy, an eval suite, or an existing agent that underperforms (flaky, wrong tool calls, retry loops, burning tokens) — even when the request just says "add a tool" or "make the agent do X" without naming a harness. Provider-neutral (OpenAI, Anthropic, MCP). Not for ordinary app features that don't change agent behavior, tool execution, context, permissions, or validation.