The Project Startup Experiment

PHAIN for short

A short exploration of how AI can be used to support the writing of an essay.

Status: 🔸 Seed → 🔸 Growing → ✅ Well-formed → 🔸 Fruitful → 🔸 Retired

Background

See The Problem with Human-AI Interaction

Approach

Typically, a project goes through a startup process to mobilise for project initiation, the first full stage of work. Other stages of work then follow. The exploration being reported here was limited to the startup process.

We used a bootstrap prompt to give the AI preliminary information about the project so that it could help us create an AI Startup Pack.

The AI Startup Pack contained the information needed by the AI for it to support our work during the project startup process.

Having installed the AI Startup Pack, we carried out part of the project initiation stage.

Process

A bootstrap prompt

The starting point was to configure ChatGPT so that it could begin to help us with this work. Goal-Directed Context Management v1#AI artefacts suggests we should have an AI startup pack to configure the AI for its initial role at the outset of a project. To obtain that, I began with a bootstrap prompt. This gave the AI an explanation of the work we were about to do together, and tasked it to help me create an improved version of the prompt.

Your first task is to advise how this prompt can be improved to make most effective use of your capabilities. But before beginning, review the information you now have in your context and ask me for any clarification you need. Do nothing else yet.

Once I had provided the clarification it sought, the AI responded with this startup prompt. It was a useful step forward, but I realised that what I really needed was a set of project instructions; I would be setting up a ChatGPT project for the essay work, and would need to manage the AI's behaviour in that space.

AI startup pack

After some discussion with the AI, we arrived at this structure for project instructions matched to the task in hand. Before turning these into a draft set of instruction, I thought I'd better remind the AI about the distinction between project instructions and project files. I asked:

Help me understand the division or boundary between the content of Project Instructions and the content of reference material in Project Files.

It responded with this explanation of the distinction between Project Instructions and Project Files. With this in its memory, I told the AI to draft a set of project instructions, using the structure we had agreed earlier, while remembering to keep what would become project files clearly ring-fenced as reference material. After some editing, the outcome was these Project Instructions for AI & Young People Essay. Together with some reference files, these constituted the AI Startup Pack.

Outputs required from startup

With the AI Startup Pack now installed in the AI's context (in its project instructions and project files), the next question was: what outputs do we need from the startup process of this project? They might include:

a clear description of the project deliverable (the essay itself);
a rationale / justification for undertaking the project;
the intended management structure for the project;
the project approach.

Typically, we would package the above in a Project Brief. We would also produce an Initiation Stage Plan to define the work needed to produce Initiation Stage products, for example the Project Plan.

In a conventional project, we would then begin the Initiation Stage. In an AI-assisted project, however, we must first produce the Initiation Stage AI Pack. This sets out the AI's configuration for that stage; i.e. it defines the context needed by the AI in order for it to support the Initiation Stage most effectively.

To summarise: after producing all the Startup products needed by a conventional project to begin its Initiation Stage, an AI-assisted project must also produce an Initiation Stage AI Pack.

The Project Brief

Task for the AI: given what's in your context, draft a clear description of the project deliverable.
Response: initially very quick but muddled, with some nonsense. I told it to think harder, and after 52 seconds it produced a much better effort.
Outcome: a comprehensive first draft of Product Description - Essay
Status: needs review

Task for the AI: draft a rationale / justification for the project
Response: a good draft
Outcome: well-written first drafts of the Rationale for the project
Status: needs review

To summarise: the AI can provide valuable, time-saving assistance for production of the Project Brief.

Findings

Similarly to when using a conventional computer, the first step was to bootstrap the AI. This involved prompting it with as good as possible a description of intentions for the project, and then working with it to develop an AI Startup Pack.

The AI Startup Pack consisted of project instructions and reference files. These were put into the AI's context within a ChatGPT project space. Working in that project space, after an initial hiccup the AI produced good quality first draft material for use in the Project Brief, one of the main outputs from project startup.

During Startup, there was quite a lot of time spent prepping the AI as compared to doing productive work with it. The balance of prep time versus productive working time is likely to shift heavily towards the latter during Initiation and later stages.

Reason for the AI's initial hiccup: Unless otherwise directed, ChatGPT-5 chooses a reasoning level based on its interpretation of the complexity of the prompt. It underestimated, in this case. Prompts need to be explicit about the reasoning level required of it.

Standard procedure should be to produce all conventional outputs required from a given stage in the project, and only then produce the AI Pack needed to support the the next stage of work.

Goal-Directed Context Management v1 was found to be a good approach to the use of AI within a knowledge-based project. It is recommended for further exploration and development.