All Skills

Interrogates research plans before evals or published claims, forcing explicit null hypotheses, confound checks, and data verification across four review waves.

Launches long-running evaluation batches in isolated tmux sessions with pre-flight verification, monitoring, and post-flight analysis for unattended runs.

Generates a CONTEXT.md routing file for a specified project subdirectory, following the canonical anatomy and ensuring it meets size and content constraints.

Simulates a NeurIPS/SC/ICSE-style peer review with five reviewer personas (HPC, ML, Stats, Reproducibility, Devil's Advocate) that verify every claim against actual result data before submission.

Reviews a plan from a prior session by extracting the canonical reference prompt from seed/plan-reviewer-design.md and invoking the plan-reviewer agent with it, surfacing the verdict without auto-applying changes.

Reviews CUDA, OpenMP, and OpenCL code for parallel correctness, covering data races, memory model violations, synchronization bugs, numerical precision issues, and API-specific pitfalls.

Reference guide for evaluating LLM-generated translations between CUDA and OpenMP, covering memory model mapping, kernel launch patterns, shared memory, atomics, and common failure modes.

Recommends the best skill, agent, or command for a given task by classifying intent and scanning available tools.

Iteratively improves a compiled LaTeX paper through two rounds of external LLM review, fix implementation, and recompilation.

Generates a structured post-task reflection capturing surprises, patterns, prompt improvements, and gotchas, writing it to a markdown file without modifying project rules.

Surface and clarify unknowns before prompting, log deviations during implementation, and verify understanding after, using HTML artifacts that pair prompts with outputs.

Guides test-driven development with a red-green-refactor loop, emphasizing behavior-focused tests and vertical slicing over horizontal test writing.

Writes or updates a HANDOFF.md file in the project root to capture goal, progress, and next steps for a fresh agent to continue in-progress work.

Synthesizes the current conversation and codebase context into a structured PRD and publishes it to the project issue tracker with a triage label.

Triage issues through a state machine driven by triage roles, including gathering context, recommending actions, and applying outcomes.

Turns a process description into a complete Standard Operating Procedure document with numbered steps, decision points, checklists, and troubleshooting guidance.

Provides nine skills for Nature-journal-family academic publishing, covering figure creation, prose polishing, manuscript writing, citation formatting, data availability statements, paper reading, reviewer responses, paper-to-PPT conversion, and academic search via an MCP server.

Runs a disciplined debugging loop for hard bugs and performance regressions, covering reproduction, hypothesis testing, instrumentation, fixing, and regression testing.

Builds a throwaway prototype to answer a design question, choosing between a runnable terminal app for logic/state questions or multiple UI variations switchable from one route.

Breaks a plan, spec, or PRD into independently-grabbable issues using tracer-bullet vertical slices, then publishes them to the project issue tracker.

Conducts autonomous multi-source research using a planner/executor architecture and an MCP server, with tools for deep research, quick search, report writing, and source retrieval.

Turns a spec or requirements into a detailed, bite-sized implementation plan with exact file paths, code, and test steps for engineers with no codebase context.

Runs Stanford's STORM pipeline to produce Wikipedia-quality research articles with citations, using Claude models and a search engine API.

Runs an interactive weekly review session to assess accomplishments against stated goals, identify missed items, and plan the next week with a structured report and accountability question.

Bootstraps the always-loaded rules layer for a new repository: writes a CLAUDE.md skeleton and a workflow.md model-notes rule, and optionally installs a native git pre-commit hook. Skips repos already rendered by the Loam Copier template.

Loads an implementation plan, reviews it critically, executes each task with verification, and reports completion, stopping to ask for help when blocked.

Turns messy business process descriptions into structured workflows, classifies each step for automation potential using a GREEN/YELLOW/RED framework, and produces a prioritized automation roadmap with time savings estimates.

Runs a multi-agent academic research pipeline covering systematic review, LaTeX paper writing, peer review, and end-to-end orchestration with integrity gates.

Scopes code-review blast radius and detects architectural hotspots using Tree-sitter AST analysis and GraphRAG, with an MCP server for graph-based analysis.

Creates isolated git worktrees with smart directory selection and safety verification for feature work.

Dispatch a code-reviewer subagent to catch issues before they cascade, with guidance on when to request review and how to act on feedback.

Creates and maintains persistent markdown planning files that survive session resets by hooking into Claude Code's lifecycle.

Guides implementing features and bug fixes test-first, enforcing the red-green-refactor cycle with mandatory failure verification.

Executes implementation plans by dispatching a fresh subagent per task and running two-stage reviews for spec compliance and code quality.

Guides completing development work by verifying tests, presenting structured options for merging, creating a pull request, keeping, or discarding a branch, and cleaning up worktrees.

Enforces running verification commands and confirming output before claiming work is complete, fixed, or passing, ensuring evidence precedes any success claims.