TABARC-Code

Beyond the Gauntlet Loop: Building AI Systems That Can Prove They Are Getting Better

The interesting future of prompting may not be better prompts at all. It may be temporary organisations of agents, critics, tests, memories and competing ideas, built around a problem and dismantled when the work is done.

There is a slightly odd thing happening with prompting at the moment and uts woeth following it down the rabbit hole.

We have spent years trying to find better ways of telling a language model what to do. Add context. (Medium article [https://medium.com/@mgr.tabarc/i-wrote-these-prompts-slightly-drunk-theyve-become-the-most-useful-ones-i-have-42ec53155bdf]) Add examples. Ask it to reason. Give it a role. Tell it not to be lazy. Tell it to check its own work. Tell it that its grandmother will be disappointed if the CSS still looks like Bootstrap circa 2016.

Then the Gauntlet Loop concept has appeared and is potentially quietly shifted the interesting bit somewhere else.

TABARC-Code

So you get a kind of generic, hedged response. It fills the space, but it doesn’t lock onto anything.

Same with:

“This doesn’t feel right, can you fix it?”

Fix what, exactly? Tone? Structure? Accuracy? Logic? You haven’t said. So it guesses. Sometimes well, sometimes not.

---

TABARC-Code

Just a quick one, really – but it’s been nagging at me.

Something I’ve noticed, using ChatGPT and similar tools, is how tightly they cling to structure. Not just grammar in the technical sense, but intent. They respond best when the language has a clear shape – when the request knows what it’s trying to be.

And English, for all its flexibility, works against that more often than we’d like to admit.

We’re loose with it. We hedge, we imply, we trail off mid-thought and expect the other person to fill in the gaps. That works fine with humans – we’re good at reading between the lines. But with an LLM, that looseness turns into noise.

So if I call someone a “turnip”, the meaning isn’t in the word itself. It’s carried by tone, inflection, timing, attitude. The context sits with the speaker, and the listener reconstructs it.

TABARC-Code

There’s a widening gap in how people understand AI at the moment, and it’s got very little to do with intelligence in the usual sense. It’s not about who’s clever and who isn’t. It’s about when you came across it, how you used it, and—crucially—which version of it you actually experienced.

Spend a few minutes scrolling and you can almost sort people into loose categories.

A large chunk had a go on ChatGPT sometime last year, usually the free version. They asked it a few questions, poked at it, maybe tried to trip it up a bit. It gave some decent answers, then something slightly off, then something completely wrong. A hallucination here, a strange phrasing there. Enough to raise an eyebrow.

And, that was that.

They;ve moved on, but the impression stuck. “Clever, but not reliable.” It settled into place as a kind of mental bookmark—filed under interesting experiment. Not revisited, not really questioned.

TABARC-Code

ok so AI Skills, and obsidian skills. the more i have been looking at the more i realised that nany including my self are treating them in isolation.

ok so been playing around with skills and back linking, with a way to se up casual linking and more graph linking protocols.

so for those interested.

---

Soft Back links

TABARC-Code

Dear Anthropic: Implement Persistent Skill Registry with Dewey Classification and Bidirectional Linking From: TABARC-Code - building sophisticated multi-skill systems Date: 30 April 2026 Status: Feature request with working proof of concept Scope: Claude.ai, Claude API, Claude Code

Executive Summary

Claude's skill system is powerful but architecturally incomplete. Skills exist in isolation. There's no persistent way to classify them, link them bidirectionally, weight them semantically, or tell Claude which skills should work together for a given task.

This proposal requests three capabilities that would transform how users build and deploy skill ecosystems: