PERSONAL PROJECT · USED AT DRANSAY
AI Automations, a registry of reusable automation definitions
- Period
- Jul 2026 to present
The problem it solves for an engineering manager
AI automations for recurring engineering work tend to start as one-off prompts, each with its own owner, access and idea of what it may change. The registry gives them one shape. Every automation carries a manifest that names its owner, risk tier, status, trigger, runtimes, skills and integrations, so the answer to “what is this allowed to do, and who answers for it?” is a file in a repository and not a memory.
It is used at DrAnsay: every team manages its own automations, and they run unattended from the company’s AI cluster.
How it works
The repository holds definitions only. Nothing in it runs by itself. A runtime loads the files and executes them, and the repository is written to work with Claude Code, Codex and OpenClaw without a rewrite per runtime.
It takes in four kinds of file:
- Automation folders, each with a manifest, a tuning file for projects, windows and thresholds, an operating-rules file and an unattended task prompt that must end in either a deliverable or
NO_REPLY. - A shared skill library, one folder per skill, following the open Agent Skills layout.
- Policies that several automations reference: risk tiers, write approval, deduplication and batching, and message style.
- Integration descriptions that record the purpose and least-privilege scopes of each external system, never credentials.
It produces a generated registry document, generated skill links for each runtime, and evaluation suites. Validation scripts check that generated files match their sources and that every owner and reviewer is on the roster.
The safety model is three risk tiers. Tier 1 is read-only. Tier 2 writes to internal systems and needs a human to approve each write until the automation’s regression evaluations pass at an agreed threshold. Tier 3 touches production or external systems and needs a security review first. The repository’s own check rejects a promotion to direct writes that lacks recorded evidence, or whose approver is the owner.
Each automation has promptfoo evaluation suites. A suite inlines the prompt, policies and skills, and a fixture supplies every tool result, so no evaluation performs an external action. Shared graders check the report envelope and delivery limits. Continuous integration selects the dependent suites when a shared policy, skill or template changes.
A registry snapshot from September 2026 lists eight definitions, all marked experimental and manually triggered at that point. They cover bug triage, test reliability, regression analysis and knowledge audits. The scripts are Python; the repository shows no other application stack.
What I learned
Start by hand. I ran the first stage manually: a person read each report before it was posted, and writes stayed behind approval until an automation’s evaluations passed.
Keep unattended work apart from interactive work. Automations that run without a person watching live in this registry. The role workflows people use interactively live in a different repository.