Weave Integration Skill

AI coding agent skill for integrating W&B Weave into existing applications — trace-first instrumentation with minimal code changes.

What It Does

Analyzes an existing codebase, creates a collaborative integration plan, and applies Weave tracing with user approval at every step.

• Scan for all functions, LLM calls, model classes, and frameworks
• Create weave_instrumentation.md with checkbox-based trace map
• AI pre-checks recommended targets, user opts out of what they don't need
• Two explicit approval gates before any code is modified

How It Works

Phase 0: Project Setup

• Confirm project name, environment, API key

Phase 1: Codebase Analysis

• codebase-analysis.md

1. Scan for ALL functions, classes, API calls
2. Detect frameworks (FastAPI, LangChain, CrewAI, etc.)
3. Check if Weave is already partially integrated

Phase 2: Plan File ⏸ APPROVAL GATE 1

• weave-instrumentation-template.md

1. Ask: "Create weave_instrumentation.md?"
2. List ALL traceable targets with checkboxes
3. AI pre-checks [x] recommended, leaves [ ] others
4. User reviews and unchecks what they don't want

Phase 3: Review ⏸ APPROVAL GATE 2

• Ask: "Approve this plan for implementation?"
• Do NOT proceed without explicit approval

Phase 4: Apply Changes

• official docs + transformation-patterns.md

1. Add weave.init() at the app entrypoint
2. Decorate checked functions with @weave.op()
3. Convert checked classes to weave.Model

Phase 5: Validation

• integration-checklist.md

1. Run integration checklist (setup, tracing, models, eval)
2. Guide user through test run + Weave UI verification
3. Ask if weave_instrumentation.md should be kept or deleted

Result

Key principles

• Trace first — list everything, user opts out
• Two approval gates — plan creation and implementation are separate approvals
• No code changes without explicit approval

Skill Structure

skills/weave-integration/
├── SKILL.md                                # Integration workflow, docs lookup, rules
├── evals/
│   └── evals.json                          # Test cases for skill trigger evaluation
└── references/
    ├── codebase-analysis.md                # Scan guide, evolve vs fresh logic, checkbox criteria
    ├── weave-instrumentation-template.md   # Template for weave_instrumentation.md plan file
    ├── transformation-patterns.md          # Model conversion patterns (class → weave.Model)
    └── integration-checklist.md            # Post-integration validation checklist

Prerequisites

• Python 3.8+
• pip install weave
• wandb login or set WANDB_API_KEY environment variable

Installation

# Clone the repo
git clone https://github.com/wandb/weave-integration-skills.git
cd weave-integration-skills

# Claude Code
cp -r skills/weave-integration/ ~/.claude/skills/weave-integration/

# Cursor
cp -r skills/weave-integration/ ~/.cursor/skills/weave-integration/

# Project-level (any agent)
cp -r skills/ .claude/skills/

Example Prompts

• "Add Weave to my existing app"
• "Instrument my FastAPI app with Weave"
• "Integrate Weave into my LangChain project"
• "Add Weave to my OpenAI Agents SDK project"
• "Retrofit Weave tracing into my production chatbot"

Documentation

• W&B Weave — Product overview
• Tracing — LLM app tracing
• Evaluation — Evaluation pipelines
• Integrations — Provider & framework integrations

License

MIT