Make AI-assisted delivery reliable and measurable.
This is where I, Eric Caskey, publish how I do that: case studies, a lesson series, a lab of working experiments, and the open-source projects behind them. It is a working notebook, not a company.
How the method works
Start with a written spec and acceptance checks. Give the agent the context for the task in front of it. Put evaluation, review, and release checks in the delivery path, where skipping them fails the build. Measure before claiming an improvement, and publish the result when it is negative.
Read lesson 1The method is public. Ballast is a RAG system that declines to answer without evidence and fails its build when evaluation scores regress. Loop Harness holds an agent to one backlog item at a time, with verification before merge. SpecSelf applies the same spec-first discipline to plain Markdown on your own machine. Read the case studies, then read the source.
Explore open-source projectsThe method applied to the platform that runs it: versioned specs, an AI agent fleet, and an attested merge gate.
Context observations from the same fleet: scoped spec slices for agents instead of documentation dumps.
A separate evidence kind: a retrieval application whose refusals and faithfulness are scored by a committed eval gate.
Start here
- Evaluate the platform work
Employer work: ownership, tradeoffs, and safety checks in a production platform.
- Learn the method
Start the engineering team of one lesson series.
- Try an experiment
Public demos, visualizations, and games.
What this site is
caskeycoding.com is Eric Caskey's working notebook on AI-assisted software delivery: what the agent is told, how its output is reviewed, what gates a release, and how the result is measured.
Case studies document personal tools and production engineering experience. Essays and lessons explain the spec-driven method. The Lab holds working software and experiments. The open-source projects, including Ballast, Loop Harness, and SpecSelf, let you inspect parts of the method directly.
Eric is a Senior Software Engineer at Amazon. This site and its projects are personal work, separate from that role. It is not a company and sells nothing.
How the work gets done
Every feature starts as a written spec in a versioned repository. AI coding agents implement against the spec, and a local-first CI gate runs the tests, the disclosure scan, and the voice check before a push is allowed. The method is documented in the lesson series and in the essays on institutional memory and model routing, and it is the same method used to build this site.
The finance material follows one more rule. The engine grades companies on quality, valuation, momentum, and health from public data, publishes its methodology, and keeps a pre-registered record of its forecasts so they can be scored later. It is an engineering exercise in measuring your own edge honestly. Nothing here is investment advice.
Common questions
- Who is Eric Caskey?
- Eric Caskey is a Senior Software Engineer at Amazon. He owns a multi-region workflow orchestration platform and an infrastructure monitoring platform. He worked at Prudential Financial from 2013 to 2022 on enterprise monitoring, reporting, and an MFA self-service platform. He writes at caskeycoding.com and ericcaskey.com.
- What is caskeycoding.com?
- caskeycoding.com is Eric Caskey's personal site for case studies, lessons, and open-source projects about AI reliability: guardrails, evaluation gates, and spec-driven agent workflows. It is not a business and offers no paid services.
- What is spec-driven development?
- Spec-driven development means every feature begins as a written specification in a versioned repository, and code, tests, and AI agents implement against that spec. On this site the specs live in a separate repository and a local CI gate checks every change against them before it can be pushed.
- Is the finance content investment advice?
- No. The finance tools and essays are educational engineering: a public factor engine, its published methodology, and pre-registered forecasts scored against a benchmark. Nothing on the site is a recommendation to buy or sell anything.