Intent & specs 1 of 6 · Intent & specs
How is work described to AI before anything gets built?
Developers prompt ad hoc, in their own words. Teams share a few prompt tips, but requirements stay in tickets and chats. Requirements are written so both humans and AI can act on them. Specs are structured and versioned, and agents draft plans from them. Specs are the primary artefact: agents build and test straight from them.
Harness & context 2 of 6 · Harness & context
What shared context and guardrails do AI tools work inside?
None. Each developer configures their own tool. Approved tools exist, but no shared conventions. Repos carry shared context: conventions, architecture notes, do/don't rules. A managed harness supplies context, tools and permissions across teams. The harness is a product: versioned, measured and improved continuously.
Pipeline & automation 3 of 6 · Pipeline & automation
Where does AI sit in build, test, review and release?
Only in the editor, as autocomplete or chat. Used for isolated tasks such as tests or documentation. AI-assisted review and test generation run inside CI. Agents take tickets to pull requests, with automated gates around them. The whole lifecycle is a loop with AI at each stage and humans at the decision points.
Roles & human-in-the-loop 4 of 6 · Roles & human-in-the-loop
How do people and AI share the work and the accountability?
Roles are unchanged. AI use is informal. Individuals experiment. Review habits are unchanged. Review expectations for AI-generated work are explicit. Roles are redesigned around builders who direct and verify agents. Small builder teams own outcomes end to end, with clear human sign-off points.
Governance & security 5 of 6 · Governance & security
How are risk, data and compliance handled around AI-built software?
No policy. Usage is invisible. A basic usage policy exists, with little enforcement. Data classification, approved models and logging are in place. Controls are built into the pipeline: provenance, secrets and dependency checks, audit trails. Governance is continuous and measured, and speeds delivery instead of slowing it.
Measurement & economics 6 of 6 · Measurement & economics
How do you know AI is improving delivery, and at what cost?
No measurement. Opinions only. Licence counts and anecdotes. Basic delivery metrics are tracked before and after AI. Cycle time, quality and cost per outcome are tracked per team. Outcome and token economics steer investment and staffing decisions.