I haven't done it with Claude Code itself, but I've built a deterministic + inferential engine that works inside a defined discipline with defined purposes. Broad enough to be practical, niche enough for most of the world to look at me funny.
The short version: let logic decide, and if there are multiple solutions let the model reason within a fixed range grounded by GraphRAG. Test every output against the logic, re-parse on contradiction, and emit 'unsure' after a couple of iterations rather than guessing.
It's no use for general knowledge. But where the judgement is largely codifiable it holds up well. There's an edge case out there that'll turn it to custard, I just haven't found it yet.
I've connected Claude Desktop to it over MCP and the results are good, not great. I designed the thing so I'm working in the sweet spot and there's still the occasional WTF.
This is very much a shared pain, though my experience is mostly outside the programming world.
I see it as a fixable pattern, but the inertia is strong.
- Lessons learned and retrospectives don't take place as often as they should
- The sessions themselves sometimes get a bit defensive, so the real nuggets get missed
- Even if a good session takes place, it ends up being static data buried in another archive
Taking an example from the construction world, I saw project after project suffer from poor planning and design. Common enough issues to warrant a moment of investigation, but not, then later being passed off as "black swan events". Because who would have thought the ground conditions would be suboptimal near a swamp?
There's something to take from the Health & Safety world. Talk about stuff. An organisation I worked for had a rule that every meeting, irrespective of purpose, had to start with a safety conversation. Something new, a reminder, a good story - just something safety focussed - and a reminder to look in the H&S system for updates and other stories. There's strong correlation between proactive conversations and reduction in safety incidents.
My thinking is that this approach should work across domains. I've done it with teams in the non-programming space with some success. A few minute knowledge-share at the start of stand-ups and regular meetings seemed to reduce the number of times known issues/challenges happened a third time. Just be prepared to lead the conversation with "I was looking through old lessons learned and saw <thing>... good reminder to <action>..." or something.
Would be awesome for industrial processes if it can be scaled, is comparable in price to heating water, and preserves what makes a good espresso a good espresso.
Was talking with a roaster who was providing espresso to a distillery recently. The distiller had tried a range of other products but only espresso shots were giving the flavour they were chasing. Needless to say, it ended up being a pretty limited run because the guys grew tired of pulling litres of shots for a batch!
Correct. An espresso drink may have milk in it. An espresso does not.
Edited to remove because the paper clarifies that it’s black coffee. But missed the (edit- found the clarification). Was quoted whilst in the process of fixing!
Something I'm seeing is good organisations doing good things, bad organisations getting worse and average organisations mostly staying flat or trending down. This is happening across multiple domains, not just an IT-centric thing.
What I'm putting it down to is the general culture/strategy of the organisation as opposed to the tools. It just so happens that tech is enabling you to accelerate whatever trajectory you were already on.
Note: This is my observation, working mostly in supply chains, at the bottom end of the earth. YMMV.
I've effectively automated about 70% of the data gathering and insight generation I need as a procurement and supply chain consultant.
I built a tool to manage survey collection and analysis of the underlying data. In the process of doing so, I stumbled across GraphRAG and Answer Set Programming. Fair to say I fell down a rabbit hole and things grew a bit...
- Category strategy and market analysis
- Supplier risk monitoring
- Emissions reporting
- Modern slavery risk analysis & reporting
I also had Claude Code build me frogger so I can properly procrastinate.
The short version: let logic decide, and if there are multiple solutions let the model reason within a fixed range grounded by GraphRAG. Test every output against the logic, re-parse on contradiction, and emit 'unsure' after a couple of iterations rather than guessing.
It's no use for general knowledge. But where the judgement is largely codifiable it holds up well. There's an edge case out there that'll turn it to custard, I just haven't found it yet.
I've connected Claude Desktop to it over MCP and the results are good, not great. I designed the thing so I'm working in the sweet spot and there's still the occasional WTF.