Experiment · EXP-004

Does re-explaining the project every turn work?

Repeated briefing versus persistent rules, files, and task-scoped chats.

In Claude and Cursor sessions, the default behaviour I saw — including my own — was to restate architecture, paste errors, and remind the model of rules. That wasted tokens and still produced amnesia. Persistent conventions and task-scoped threads worked better in practice.

Hypothesis: If I explain the project thoroughly enough in the chat, the model will keep it.

Problem

A common workflow with coding assistants is to repeatedly explain the same architecture, paste the same conventions, restate the same constraints, and reintroduce the same errors every time a new task begins. That is not primarily a model problem. That is a context problem.

A prompt is only one part of model input. In production, context is assembled from the request, previous interaction, retrieved files, application state, tool outputs, and system instructions. The engineering problem is the smallest, most relevant, most trustworthy collection of information the model needs at this moment.

Hypothesis tested

I initially treated a sufficiently detailed chat briefing as a substitute for project memory. In practice, repeated briefing consumed tokens, made conversations longer, and still allowed important rules to disappear inside unrelated context.

Alternative workflow

I moved durable conventions into project rules and files, kept architecture information close to the codebase, asked the agent to read only relevant files, and separated unrelated tasks into different conversations. Errors became task inputs rather than permanent transcript history.

Result

The workflow became more predictable because persistent information stopped depending on whether I remembered to restate it. This also reduced the amount of duplicate context sent on every turn.

Finding

Briefing is not infrastructure.

If information should remain true across tasks, it belongs in a persistent project artefact or an explicit memory layer, not only inside a long chat.

Abubakar, Imam. "Does re-explaining the project every turn work?." Imam Abubakar AI Research, 2026. https://imamabubakar.com/research/exp-004-repeat-the-project

This archive contains independent research, experiments, architecture studies, and technical observations. Some findings are exploratory and should not be interpreted as peer-reviewed scientific conclusions unless explicitly stated. Architecture studies of third-party products are based on publicly observable behaviour, published technical information, and engineering analysis. They are not claims of access to private source code or infrastructure.