Skip to content

Repository files navigation

SciAi Skills

Custom Claude Code skills for scientific computing workflows. These skills provide multi-agent orchestration for research, plan creation, review, splitting, code verification, and adversarial GPT critique.

See the Skills table below for available commands.

Skills

Skill Command Description
sciai-verify /sciai-verify Multi-level correctness verification with independent AI reviewers (Sonnet/Opus/Codex)
sciai-plan-review /sciai-plan-review Multi-agent plan review for completeness, feasibility, and goal alignment
sciai-plan-split /sciai-plan-split Decompose large plans into parallel/serial sub-plans with dependency analysis
sciai-batch-review /sciai-batch-review Review all split sub-plans in parallel + integration review for cross-plan coherence
sci-planner /sci-planner Create verified, actionable plans with auto-review and iteration
sci-research /sci-research Deep parallel research using specialized web, literature, and codebase agents
gpt-critique-loop /gpt-critique-loop Iterative adversarial GPT (Codex) critique loop that hunts for holes in a plan, derivation, or hypothesis and iterates until it passes

Verification Levels

All review/verify skills support three thoroughness levels, letting you trade speed for depth:

  • Level 1 -- Sonnet only (fast, ~2 min)
  • Level 2 -- Sonnet + Opus in parallel (~5 min)
  • Level 3 -- Sonnet + Opus + Codex in parallel (~10+ min)

Typical Workflow

1. /sci-research              -> understand the landscape
2. /sci-planner               -> create a plan (auto-reviewed)
   |-- if too large           -> /sciai-plan-split -> parallel sub-plans
                              -> /sciai-batch-review -> review all + integration check
   |-- to harden the plan      -> /gpt-critique-loop -> adversarial GPT review until approved
3. [implement the plan(s)]
4. /sciai-verify 2            -> verify the implementation

Installation

Each skill directory contains a SKILL.md that defines the skill's behavior. Copy or symlink the skill directories to ~/.claude/skills/ to make them available in Claude Code.

# Example: install all skills
for dir in sciai-verify sciai-plan-review sciai-plan-split sciai-batch-review sci-planner sci-research gpt-critique-loop; do
  ln -sf "$(pwd)/$dir" ~/.claude/skills/"$dir"
done

See skills.md for detailed documentation on each skill.

License

MIT

About

Multi-agent AI workflows for scientific-computing research with Claude Code — parallel literature review, multi-model correctness verification, and adversarial GPT critique loops to plan and harden numerical methods.

Topics

Resources

Stars

1 star

Watchers

1 watching

Forks

Releases

Packages

Contributors