Overview — Project 8: Rankforge

Rankforge begins with a real TypeScript experiment: scrape a business website, research search demand, choose one angle, outline it, write the article, package it for SEO, and print estimated model cost. That experiment is useful because every run produces something you can read. It is also a perfect Rust AI engineering exercise because a single script hides several boundaries that production software must make explicit.

You will preserve the fast feedback loop while turning the workflow into typed, testable stages.

Run the finished project

From rust/rust-ai-engineering:

cargo run -p rankforge -- build \
  --source fixtures/rankforge/site.md \
  --responses fixtures/rankforge/responses.json \
  --output target/rankforge

Inspect the result:

target/rankforge/
├── blog-1-research.md
├── blog-2-angle.md
├── blog-3-outline.md
├── blog-4-article.md
├── blog-5-seo-package.md
├── blog-final.md
├── blog-cost.json
└── run-report.json

The required build is offline. responses.json acts like saved responses from a scraper, web-search tool, and model provider. This lets you learn the pipeline and reproduce failures before a network, API key, changing search result, or model release enters the system.

The five gates

Day	Product slice	Rust pressure	AI engineering lesson
35	CLI and website snapshot	paths, bytes, ownership, errors	acquisition is separate from generation
36	five-stage workflow	enums, traits, exhaustive order, artifacts	workflows beat one giant prompt
37	search-backed research	typed tool observations, provenance	search output is untrusted evidence
38	article and SEO package	validation, strings, fail-closed writes	format checks and factual checks differ
39	receipts and completion gate	checked math, hashes, integration tests	measure cost per inspectable completion

The pipeline you own

website snapshot
      │
      ▼
research ─▶ angle ─▶ outline ─▶ article ─▶ SEO package
   │          │          │          │            │
   └──────────┴──────────┴──────────┴────────────┴──▶ stage artifacts + receipts

Each arrow is a gate. The next stage receives an accepted artifact, not an invisible conversation history. If the angle is weak, you inspect blog-2-angle.md. If packaging removes useful depth, you compare it with blog-4-article.md. If spend jumps, you inspect usage per stage.

What changes from the TypeScript experiment

Rankforge keeps the original five useful steps and improves their software boundary:

• stage order is an enum-backed contract;
• every intermediate response is a durable artifact;
• required sections are validated before publication;
• punctuation policy is executable rather than prompt-only;
• token and tool spend use checked integer arithmetic;
• every stage receipt binds its input hash to its output hash;
• the required tests do not spend money or depend on current search results.

The live extension can still use a scraping API, the OpenAI Responses API, and web search. Those are adapters at the edge. The workflow and its completion evidence remain yours.

Start with Day 35 — Freeze the Source →.