Back
Jamie Watters
Jamie Watters
@TheWayWithin
Systems programmer who loved the work, got scared, hid in management for 20 years. Now engineering my way back out with AI and Claude Code.
Swapped the signal model after replaying 288 saved production prompts offline. R:R compliance went 37% to 100%. Write-up: jamiewatters.work/journey/llm-ignoring-its-prompt 
#trader7
Built a claim-verification script that extracts every number, date and quoted phrase from a draft and fails unless each one traces to the evidence base. 1,727 claims, exit 0. 
#jamiewatters
Published: Why AI projects fail: a working paper tracing 19 headline AI failure statistics to source. Seven turn out not to be measurements of anything that happened. jamiewatters.work/journey/why… #jamiewatters
Published a new post: Soft Systems Methodology explained. Checkland's 1969 method for the AI brief six people read four different ways. jamiewatters.work/journey/sof… #jamiewatters
Published a new post: 80% of AI projects fail on the problem, not the model. RAND's leading root cause isn't the tech, it's that nobody agreed what problem was being solved, and Checkland's Soft Systems Methodology has been aimed at that since 1969. jamiewatters.work/journey/ai-… #jamiewatters
Swapped both LLM decision models from Opus 4.7 to 4.8 with one env var, validated on 78 replayed prompts first. Didn't miss a beat. #trader7
Worked out that the interface is lagging the capability: the models got better at doing, nobody worked out how they should talk about it
Wrote up why Opus 5 makes me swear, and admitted that half of it is my own bandwidth rather than the model