AI writes plausible code fast, and every new session is a stranger to the last. Generative Specification keeps intent explicit enough that any reader, human or AI, can rebuild the system without guessing.
AI writes plausible code fast, and then every new session is a stranger to the last. Generative Specification keeps intent explicit enough for a stateless reader, human or AI, to rebuild a system without guessing. A specification, a harness that verifies it against a running system, and a standard you can hold a codebase to. Everything here is designed to be used, not just read.
Whether you write the code or you own it, the fastest path is to see the discipline work on a real codebase.
Try it on your own code in 20 minutes, no install. Score a repo, see the failure modes named, get a remediation plan, then use the gates and spec templates to hold the line across sessions.
Get an independent read you can act on: a Readiness Assessment scored to a maturity level, remediation that installs the guardrails in place, or technical due diligence with a clear asset / conditional / liability verdict.
"Everyone feels like a systems engineer capable of building things now, but the truth is it gets us into trouble, especially at publish time."
"A non-technical analyst on the team built an audit agent for their CRM from what he learned, documented it, and presented it to leadership. They called it complete."
A programming paradigm for the stateless reader. Seven properties that make a specification derivable by an AI that carries no accumulated context, so the decisions that hold a system together stop being forgotten between sessions.
Structured programming operates at the syntactic tier. SOLID, TDD, and DDD operate at the semantic tier. Generative Specification operates at the pragmatic tier, where derivability by a stateless reader becomes a binding constraint. The full argument, the formal lineage from Aristotle to today, and the empirical evidence. Open access on Zenodo.
Free tools for individual developers. Each one addresses a specific failure mode of AI-assisted development.
Portable quality gates, one per property, wired to standard tools: mutation, real coverage, complexity, duplication, dead code. Hand it to your AI and it cables the gates into your project straight from the spec. The ratchet only goes up.
Quality gates →Graph-powered hybrid search for your codebase. Vector + BM25 + path, fused with RRF. Semantic search, dependency traversal, and contextual file reads — the AI reads your codebase the way you do, not just grep.
GitHubThree-tier persistent memory for AI sessions. Buffer, Working, and Core layers with Ebbinghaus decay and an intelligence distillation layer. Cross-project knowledge that survives context boundaries.
GitHubTwo days. Your team. Your codebase. A different speed.
Point AI at a codebase and you get plausible output fast, then pay for it in review and rework. The Forge teaches the discipline that makes the output correct enough to trust: spec-driven development, clean architecture, and directing AI against a spec instead of prompting into the dark. Two days, on your actual codebase, with your team.
Book a ConversationLong-form writing on specification, AI-assisted development, and what it means to engineer in an age of stateless readers.
ambientengineer.dev The Harness Manifesto →Workshop inquiries, research questions, collaboration.