Acceptance first. Agents second.
ADAC Skills turns AI coding from plausible patch generation into evidence-backed engineering delivery. Define context, boundaries, user-verifiable acceptance checks, and human gates before implementation expands.
A delivery contract for AI coding agents.
ADAC is broader than TDD and stricter than prompt guidelines. It asks agents to produce reviewable acceptance evidence and user-runnable scenarios, not just code that compiles.
Context
Make the system, business rules, history, and known risks available before generation.
Spec
Define target behavior, non-goals, boundaries, and fallback before implementation.
Evals
Turn real risks into behavior checks, user verification scenarios, regression, performance, lifecycle, and observability evidence.
Gates
Reserve architecture, concurrency, security, release, and residual-risk calls for people.
Where ADAC helps.
Use it when agent output is fast but correctness depends on engineering judgment, historical regressions, or release safety.
Async resource loading
Before code, ADAC asks for user-runnable scene checks, lifecycle evidence, same-scene timing, logs, telemetry, and concurrency gates.
Checkout coupon path
ADAC classifies payment changes upward, then requires pricing invariants, audit evidence, and security/product gates.
Incident workflow skill
ADAC turns one-off incident fixes into reusable skills with evidence collection and approval boundaries.
Install and validate.
Install the portable ADAC skill with the skills CLI, then validate ADAC plans with the npm command.
# portable skills CLI npx skills add HaloForgeAI/adac-skills --skill adac # local validation npx skills add . --list
# plan validation npx adac-skills validate examples/worker-module.adac.md # package metadata npm view adac-skills