Guide · Directing agents

Delegate the goal, not the task

That an agent takes a goal where a chatbot takes a task is the easy part to say — and the shift itself has its own guide. This one is about doing it well: a framework for briefing an agent, and the move that makes a good brief almost effortless — building it into your setup so you stop re-typing it.

The practical difference

What you hand over changes.

With a chatbot you direct a task: you spell out the role, the task, the format and the context, and it does that one bounded thing. With an agent you delegate a goal: you give it the outcome, the rules it must stay inside, and how to judge its own work — then let it choose the steps. Micro-management becomes macro-management; operator of a tool becomes manager of a colleague.

Which means the useful question stops being “what is the perfect prompt?” and becomes “what does a good brief always contain?” Answer that once and you can give it every time. A good brief has five parts.

The framework

CARD — the five parts of a brief.

The format I teach for directing an agent: Context, Artefact, References, Destination — and the quiet fifth, Validation.

C

Context

What the AI needs to know about the situation — the background it reasons from.

A

Artefact

What it should produce — the specific thing you want back.

R

References

What to follow — the sources, methods and standards the work must respect.

D

Destination

Where the output goes, and in what form — the file, the format, the reader.

+V

Validation

How both of you will know it is good — the self-check it runs before the work reaches you.

The unlock

Stop re-typing the brief. Build it into your setup.

Here is the part that makes this practical rather than just tidy. Four of CARD’s five parts barely change from one job to the next. The references you follow, where a file belongs, the standard it is held to — those are standing facts about how you work. Re-typing them into every prompt is exactly the micro-management you were trying to escape.

So don’t. Give each standing part a home in your environment, and the AI reads it before you say a word.

Context Your working files. The project or client folder you start from — its background is already there to read.
References Your knowledge base & instruction files. Your methods, house standards and reusable skills — so it follows your way of working, not generic training data.
Destination Your naming & routing rules. Where a file of this kind belongs, what it is called, and where it goes — already decided.
Validation Your standards rubric. What good work looks like here — the bar, written down once.

Two of those homes are the plain instruction files a practice runs on — the standing orders and the reusable skills. They get their own guide: the files that run an AI practice.

And take Destination, the part that varies most. A finished thing rarely just sits in a folder. In my own practice the same output often has several destinations at once: a guide like this one is committed to a code repository and published to a website in one motion; a client note is a document that leaves by email; another piece is dropped in a shared drive or pasted into a chat tool. Each wants a different shape — Markdown for the repo, a clean document for the inbox, a web page for the site — so a setup that knows where a thing is headed produces it in the right form and routes it there, instead of leaving you to convert and file it by hand.

The better your environment, the less you have to say. The structure becomes the standing brief.

In practice

One delegation, before and after.

Say I want a proposal drafted for a prospect. Here is the same delegation without a built setup, and with one.

Without the setup

A paragraph of instructions

“Draft a proposal for this prospect. We’re a training and AI-governance consultancy; match our house tone — plain, answer-first, no hype. Use our standard pricing tiers. Lead with the problem in bold. Put it in our proposal format, save it where proposals live, and don’t put fees on the first page…”

Re-typed, more or less, every single time.

With the setup

A goal

“Draft the proposal for this prospect.”

The setup already holds the tone, the pricing, the format, the destination and the bar. I supplied the only thing that was new: the goal.

That is the whole thesis in one example. A well-built environment doesn’t just store your files — it holds your standing direction, so a real delegation collapses to its variable core. You hand over a goal, not a specification.

Where it fits

Direction is one of four competencies.

CARD is the Direction competency in its agentic mode. It sits inside a loop of four — our adaptation of the AI-fluency framework (Anthropic / Dakan / Feller).

Before

Delegation

Deciding what to hand the AI at all, and at what intensity. Stays human.

The brief

Direction

CARD (+V) — communicating the goal so it can act. Increasingly carried by your setup.

After

Evaluation

You checking the draft against the bar — facts, logic, tone, fit. Where your judgement goes.

The signature

Ownership

Your name on the result. Never “the AI did it.” Stays human.

Notice where the work moves. Delegation and Ownership stay firmly human — what you choose to hand over, and whose name is on the result. Direction, and the standard built into it, get carried more and more by the setup. Which leaves Evaluation — your read-through against the bar — as where your judgement is actually spent. Standard, check, signature.

From one desk to many

The same architecture, scaled.

I run a solo practice this way — but nothing about it is solo-only. The four homes that make CARD physical on one desk are the same four things an organisation has to get right to use AI well, scaled up:

Context becomes governed, shared data — the material teams are cleared to reason over.

References become a shared knowledge base and an approved set of tools and skills — one house method, not forty private ones.

Destination becomes where outputs are allowed to land and publish — the routing, and the record of what was produced.

Validation becomes the standard the organisation holds AI work to — the rubric, applied before anything ships.

Read that list again and you are reading AI governance, made operational. A sound personal AI setup and a sound organisational one are the same four anchors at different scale — which is also why governance can’t just be a policy in a binder. It has to live in the setup. That is the harder question of who owns it.

Learn to delegate to AI — and build the setup that carries it.

Kramer Consulting runs hands-on sessions on directing agentic AI — CARD, the four competencies, and putting the brief into your own environment — on your real work, with your people owning the result.