Tool Building Workflow

conceptdevelopingCreated: 2026-07-25Updated: 2026-07-25
blog-empireworkflowtool-buildingai-toolsreference

Tool Building Workflow

The tool building workflow is the pipeline for creating ToolBrain tool pages. It follows the repo-prompt pattern — analyzing a GitHub repository, constructing an evaluation prompt, running it through multiple models, and publishing the results as a structured tool review.

The Repo-Prompt Pattern

The core workflow for each tool page:

Phase 1: Repository Analysis

  1. Select a tool/package from the tool queue or discovery pipeline
  2. Clone or fetch the GitHub repository
  3. Analyze the repo structure: README, documentation, code samples, API surface
  4. Extract key claims, features, version info, and usage patterns

Phase 2: Prompt Construction

  1. Build a structured prompt incorporating repo analysis
  2. Include evaluation criteria (correctness, UI/UX, code quality, pattern adherence)
  3. Specify output format expectations
  4. Include the HERO_IMAGE_PROMPT directive for generated hero images

Phase 3: Model Evaluation (Arena Mode)

  1. Run the prompt through multiple builder models
  2. Judge models score the outputs using the Arena Workflow
  3. Determine winner based on aggregate scores
  4. Extract hero image prompt from winner

Phase 4: Publication

  1. Strip meta-instructions from output (HERO_IMAGE_PROMPT lines, etc.)
  2. Add frontmatter (title, slug, tags, publish date)
  3. Pass through quality ratchet (sourcing, claims audit, image verify, scoring)
  4. Publish to ToolBrain via Ghost CMS

ToolBench Specification

A formal benchmark for evaluating AI models building AI tools, defined with 4 dimensions:

Dimension Weight
Correctness 0.35
UI/UX 0.25
Code Quality 0.20
Pattern Adherence 0.20

Five benchmark problems:

  1. Token Counter
  2. Prompt Analyzer
  3. Model Compare
  4. Version Diff
  5. Template Builder

Quality Considerations

  • Claim audit: Every claim in a tool page must be sourced. The claim-audit.py script fetches cited URLs and verifies claims match sources.
  • Citation rate target: >80% (ToolBrain is currently at 47%, flagged as failing)
  • Hero images: Feature image rate target >90%

Related