What's working for us with spec-driven development is treating acceptance criteria as checks the agent can run, not paragraphs it has to interpret. Small TOC example in the latest post on the prickles blog.
If the acceptance criteria cannot be checked, the agent is still guessing. Turning specs into runnable checks gives the agent a feedback loop instead of just more text to interpret.
Suny Choudhary
Building AI Security for LLMs | CEO @ LangProtect
This makes sense.
If the acceptance criteria cannot be checked, the agent is still guessing. Turning specs into runnable checks gives the agent a feedback loop instead of just more text to interpret.