Since April 2026, I’ve been building a series of production-grade tools on the Anthropic Claude API — using Claude Code as the development environment. Each project is scoped to roughly a day of active work and is designed to demonstrate a specific set of API capabilities in a real, useful application.
This isn’t “hello world with an LLM.” Every project ships with a full test suite, adversarial edge cases, and documented build decisions. The goal is to show what’s possible when you actually learn the tools — not just know they exist.
Projects build on each other. The API features compound.
| 01 | Legacy Code Explainer (Python + Claude API) Point it at a legacy source file — PHP, Perl, COBOL, C, Tcl/Tk — and it streams back a structured risk report: what the code does, what’s dangerous, and how to modernize it. Built for the “this has been running since 2003, and nobody knows why” problem. streaming | prompt caching | extended thinking | Python · CLI | pytest · 24 tests |
| 02 | PDF Contract Analyzer (Django + React + Claude API) Upload any contract PDF — NDA, SaaS agreement, employment contract, commercial lease — and get back a severity-bucketed risk report with every dangerous clause flagged and cited with exact verbatim text. Businesses pay $300/hr for this. It takes about 30 seconds. PDF support | extended thinking | structured output | citations | Django · React · Tailwind | 94 tests |
| 03 | Changelog & Release Notes Generator (FastAPI + Claude API) Paste a git log, choose your audience — developer, client, or executive — and get back polished, structured Markdown ready to drop into GitHub Releases, a client email, or a Notion page. Same commits. Three completely different documents. prompt engineering | audience modes | input resilience | FastAPI · Python · Pydantic | pytest · 55 tests |
| 04 | Next project — in development coming soon Each project in the series adds a new API capability on top of the last. Check back soon, or follow along on GitHub. |
Feasibility Studies
Not every build is scoped to ship. Some are scoped to answer a question — and a clear “no, and here’s exactly why” is as complete a result as a working app, if it’s documented with the same rigor. These don’t carry numbers in the main series; they’re not building toward the compounding set of API capabilities the numbered projects are. They’re standalone tests of a specific approach, kept here because the finding is worth as much as a shipped tool.
| FS-01 | Daily Job Scrub Pipeline (Python + Claude Cowork) Question: can an agentic Claude Cowork session reliably drive a daily Greenhouse job-scraping pipeline at scale? Built the full thing — fetch, dedupe, filter, rank, AI-assisted company vetting, AI-assisted resume fit scoring — over three days. The Python core works and is solid. The agentic session never completed one clean end-to-end run, at a scale small enough that a post-hoc cost estimate ruled out token budget as the cause. Answer: no, not as architected — and the four specific failure modes that proved it are documented in full, along with the corrected design for round two. retrieval-augmented synthesis | structured comparison | agentic-session failure modes | Python · openpyxl · Claude Cowork |
| FS-02 | Next project — in development coming soon Future feasibility studies will test other approaches before they get built out fully. Check back soon, or follow along on GitHub. |
How the Series Works
Every numbered project in this series follows the same constraints. (Feasibility Studies, below, are a separate track — same honesty standard, different goal: testing whether an approach works at all, not shipping a polished tool.)
- Scoped to one day — approximately 2–3 hours of active development with Claude Code across 10–15 commits. The
SESSION_LOG.mdin each repo is a timestamped build diary documenting every phase, including wrong turns. - Production test coverage — each project ships with a full pytest and/or Vitest suite, mocked API calls, adversarial fixtures, and documented edge cases. No “it works on my machine” code.
- One new API capability per project — streaming → prompt caching → extended thinking → PDF support → structured output → citations. The capabilities compound.
- Real problems, not demos — each tool is something a real business would pay for. The benchmark isn’t “does it call the API” — it’s “would someone actually use this.”
The AI-assisted build process is part of the story. The session logs are committed intentionally.
API Capabilities Covered Across the Series
| Capability | First introduced | What it enables |
|---|---|---|
streaming | Project 01 | Token-by-token output; no spinner, no wall of text on completion |
prompt caching | Project 01 | 3,000+ token rulesets served from cache on repeat runs — sustainable cost model at volume |
extended thinking | Projects 01 & 02 | Claude reasons through ambiguous clauses before producing output — visible quality difference on genuinely complex inputs |
PDF support | Project 02 | Documents sent as native base64 blocks — no text extraction, no preprocessing, no lost formatting |
structured output | Project 02 | Responses constrained to a strict JSON schema — every flag has severity, category, explanation, and verbatim citation |
prompt engineering | Project 03 | Modular audience injection — persona, tone, and format as independently tunable blocks |
Background
Seven years at CalPERS building internal workflow automation in PHP, JavaScript, and MySQL means I’ve watched good teams ship great software — and then write terrible changelogs, miss risky contract language, and maintain codebases nobody fully understands anymore. These tools address real problems I’ve seen in production environments, not hypothetical ones.
Source code for every project is on GitHub. You’ll need an Anthropic API key to run them locally, available at console.anthropic.com — new accounts include free credits.