I Hate Being Human Glue Between AI Steps - Spec File as the Deliverable
I Hate Being Human Glue Between AI Steps
The promise of AI coding agents is autonomy. The reality is that you spend half your time copying output from one step, reformatting it, and pasting it as input to the next step. You are the glue.
Agent says "I created the database schema." You check it. It looks fine. You tell the agent to generate the API routes. It asks what the schema looks like. You paste the schema back in. The agent generates routes. You tell it to write tests. It asks what the routes look like. You paste the routes back in.
This is not automation. This is a fancier clipboard.
The Spec File Approach
The fix is making the specification the deliverable, not the code. Instead of asking an agent to "build a user authentication system," you write a spec file that describes every component, how they connect, and what the expected behavior is.
The spec file becomes the single source of truth. Every agent step reads from the same spec. No human glue needed.
A practical spec file includes: data models with field types, API endpoints with request and response shapes, error cases and how to handle them, and integration points between components. Write it once, and every agent that touches the project can read it.
CLAUDE.md as a Spec Layer
If you use Claude Code, the CLAUDE.md file is already a natural place for specifications. Put your feature specs there. The agent reads them automatically at the start of every session. No copy-pasting required.
The pattern works well: one CLAUDE.md per feature branch, describing exactly what needs to be built. The agent reads the spec, builds the feature, and checks its work against the spec. You review the output once instead of babysitting every step.
Why This Matters
Every minute you spend being human glue is a minute you are not thinking about the product. Specification-first development is not about writing more documentation - it is about writing the right document once so agents can work without you hovering over them.
The best agent workflows are the ones where you define the destination and walk away.
Fazm is an open source macOS AI agent. Open source on GitHub.