The Art of Specification

A quote from Nate Jones, AI Communicator.

Nate made this point mid Feb 2026 in the context of the Clawdbot explosion (whose founder then got pulled into OpenAI).

In my world as a business-process (AI) automation consultant, I see exactly this all the time, both within my own work and in the deployments my clients run.

Organisations are deploying agentic AI with genuine excitement and genuine results.

But the failures?
They almost always trace back to the same root cause: a lack of definition at each layer of operation.

I use a three-layer framework in my agentic build-outs to bridge that gap
(hat tip to Nick Saraev ):

1. Directive (Intent)
> What are we trying to achieve, and why?
> The human intention up-front in a structured, SOP-like format. (If the 'why' is fuzzy, the agent's what will be a lottery).

2. Orchestrate (Logic)
> The framework to govern how the agent coordinates its work. Assuming that the AI will just figure it out is setting up for inconsistent results, not acceptable to business.

3. Execute (Action)
> Non-ambiguous, endlessly repeatable functions called at the end of the chain.
> Pure execution with zero interpretation.

[This method largely mirrors the philosophy behind Claude Code’s ‘skills’ architecture of high-level capability backed by rigid, functional execution.]

By their nature, agents will attempt to interpret your intent when put to work.
Absent clear and well-structured instructions, those interpretation will near-certainly deliver chaos over time.

It’s not a tech problem, it’s a specification issue.

In a way that’s always been our challenge - reaching back to the building of the Pyramids.

Closing the gap IS the work.