Author built acai.sh, an open-source spec-driven development toolkit using feature.yaml and numbered ACIDs to anchor AI agent output to concrete, trackable requirements.
Key Takeaways
ACIDs (Acceptance Criteria IDs) are numbered requirement tags that agents embed in code and tests, enabling acceptance coverage tracking instead of just test coverage.
feature.yaml is a structured alternative to markdown PRDs: hierarchical components, numbered requirements, and cross-references between specs.
The acai CLI (acai push) ships spec and code-ref data to a hosted dashboard for requirement-level PR review, replacing file-by-file GitHub diffs.
Hosted dashboard (Elixir/Phoenix/Postgres) is free for now; CLI is on npm and GitHub (Apache 2.0); source is open.
Author frames specs as the only durable artifact: code, tests, and prompts become disposable; acceptance criteria do not.
Hacker News Comment Review
Commenters broadly agree the underlying idea is old: V-Model, BDD/Cucumber, ADRs, and formal specification from the 1960s-70s all cover this ground; the YAML format and ACID tagging are the novel surface layer.
Skepticism surfaced around YAML as a spec language: one commenter sarcastically suggested adding Jinja, echoing a known pain point about YAML complexity creep.
The “overcame AI psychosis” framing drew pushback: the post reads as going deeper into LLM tooling, not stepping back, with the toolkit itself being an AI harness for building products.
Notable Comments
@jFriedensreich: “Where is the part where the author overcomes ai psychosis? Reads like digging in deeper and deeper.”
@stevefan1999: frames acai as Cucumber/BDD repackaged in YAML so LLMs can parse AST instead of tokenizing natural language.
@beshrkayali: agent-generated code loses institutional memory; specs become the only surviving record of intent when authorship is diffuse.
