The art of goal-directed context management

with acknowledgement to Erling S. Andersen, whose book Goal Directed Project Management inspired me in 1988

Transparency label: human only

How do we control an AI?

To use an AI effectively, we have to configure it’s behaviour to suit the purpose of our work. We do this by giving it information in the form of instructions, guidance, questions, and reference material. We put this information into the AI’s context to govern how it behaves. See What is context? for a more detailed explanation of context. It is important to understand it well.

To apply a standard behaviour from the AI across all chats, we use custom instructions.

During individual chats, we control the AI through prompts, the canvas, and file uploads (prompts can include instructions, guidance and questions).

To control the AI’s behaviour for a group of chats organised as a project, we do it through project instructions.

To make information available to all chats in a project, we upload project files.

Finally, if we need the services of an AI specialist, we create a form of agent (called a GPT). We configure the agent with bespoke instructions to make its behaviour appropriate to its specialist role, and we upload reference files to give it a knowledge base. Our agent can then be called in as necessary to provide specialist advice through chats.

This method of controlling the AI by putting information into its context is similar to the way we control a computer by giving it a program to run. But there is an important difference: a computer executes the instructions in its program precisely; an AI (large language model) interprets the information we give it. We should aim, therefore, to give the AI all the information it needs to avoid losing focus. A vague prompt gets a vague answer. A precise prompt with the right background information gets us something much more useful.

There are similarities with the way we manage a human. The human has a store of learned knowledge to call upon, can be given instructions on how to behave during a task, can use reference material, can converse with others, and can co-edit documents with them. Like an AI, the human interprets the information obtained by these means.

Managing scale & complexity

As the scale or complexity of a task grows, it becomes increasingly difficult to coordinate the work of humans to achieve a common goal. To overcome this problem, we use project and programme methodologies to structure the work; these provide a form of scaffolding to let the work be carried out safely and effectively. The scaffolding is made strong and stable through a plan which – amongst other things – divides the work into manageable packages, defines who or what will do the work, when and how it will be done, and what work outputs are required.

Work whose purpose is well defined and carries an acceptable risk of failure can be organised as a project. The scaffolding for this takes the form of the project plan and related documentation, including external standards. Members of the project team are given roles, which can vary during the project, and assigned tasks. The detailed information they need to carry out their work successfully is held in the project scaffolding. When the work has greater scope and/or complexity, it can be organised as a programme. In this case the scaffolding is provided by the programme management plan and related documentation.

AI as a team member

When AI is used by a member of a project team, its characteristics make it more like an additional member of the team than a software tool. The AI, like the human, must be given roles and assigned tasks. It, too, must obtain detailed information from the project scaffolding. Unlike a human, however, the AI cannot yet go and fetch the information it needs; the human user must provide it. The human does does so by putting relevant information into the AI’s context.

Information required for input to the AI’s context over the course of the project is held in a part of the project scaffolding called the contextual scaffolding. Although closely integrated with other parts of the project scaffolding, the contextual scaffolding has a distinct purpose: to hold the information – AI artefacts – needed to manage the behaviour and effectiveness of the AI over the course of the project. The contextual scaffolding is the responsibility of the project’s AI manager.

AI artefacts

The nature, required content, and production timetable for AI artefacts are governed by and tied to the project plan. Artefacts in the contextual scaffolding include:

AI policy
AI management plan
an AI startup pack to configure the AI at the outset of the project, when its objective will be to outline the contextual scaffolding and help to produce an initial set of AI artefacts;
stage packs tailored to the varying uses intended for the AI over the course of the project (as with other project artefacts, these begin life in outline form and gain content as the project progresses);
GPT packs to create the specialist AI agents that are needed.