datascience

stellars-claude-code-plugins

Your AI agent will cut corners. This is the forcing function.

You ask Claude to "improve error handling." Claude says "Fixed it." Two files changed, no tests run, edge cases broken. Or it ships an SVG infographic with overlapping text and contrast failures. Or it passes a document past a reviewer who'd tear it apart.

This marketplace makes Claude work like a disciplined engineer instead. Each plugin enforces a specific discipline: research before implement, validate before ship, ground every claim, audit every iteration.

# Force Claude through research, plan, test, review, and audit before claiming done
/autobuild:run improve error handling in the API layer

#  -> writes PROGRAM.md (objective + scope)
#  -> writes BENCHMARK.md (measurable score)
#  -> asks for your approval
#  -> implements
#  -> runs tests
#  -> reviews against the benchmark
#  -> records evidence in YAML audit log

/plugin marketplace add stellarshenson/claude-code-plugins
/plugin install autobuild@stellarshenson-marketplace

Read the long-form articles: Your AI Agent Will Cut Corners. Here's How to Stop It and Stop Fixing Your AI's SVGs. For real examples (60+ production SVGs, 4 worked devils-advocate analyses, 3 autobuild iteration trajectories, a 1.0-CV grounding result), see showcase/.

The full marketplace - six disciplines

autobuild is the spear. The same forcing-function logic powers five more plugins, each enforcing a different kind of discipline on Claude. Install them individually or as a bundle.

Plugin	What it solves
autobuild	Executes code and artefact builds toward an objective with iterations driven by a calculated outcome benchmark - enforces structured phases with multi-agent review
devils-advocate	Produces high-quality documents for a specific audience using a scientific, measured, iterative approach - quantified critique with Fibonacci risk scoring and per-iteration residual measurement
svg-infographics	Produces high-quality standardised SVG infographics - grid-first design, theme-driven styling, dark/light mode, 5 routing modes (straight/L/L-chamfer/spline/manifold) with A* auto-routing, callout placement solver, chart generation, and 6 automated checkers
datascience	Produces high-quality data science projects and notebooks following consistent standards - scaffolds projects from copier templates, enforces notebook structure, applies rich output styling, and supports prompt engineering techniques
document-processing	Processes documents according to user requests with grounding in source materials - source tracing, compliance checking, PDF automation
journal	Produces a work journal marking key changes, implementations, and decisions - append-only audit trail with continuous numbering, archiving, and deterministic journal-tools CLI for validation, sorting, and word-count enforcement

autobuild

Runs structured multi-iteration development cycles where each iteration passes through a full phase lifecycle with quality gates. A program defines what to build, a benchmark measures progress, and the engine enforces the workflow until the objective is met or iterations are exhausted.

• Shallow fixes - forces research and hypothesis before implementation
• Scope creep - plan locks scope, review catches deviations
• Lost context - hypothesis catalogue and failure context persist across iterations
• Unchecked quality - two independent gates (readback + gatekeeper) per phase
• No accountability - every phase records agents, outputs, and verdicts in YAML audit logs
• Benchmark gaming - guardian agent checks for benchmark-specific tuning vs genuine improvement