One structure. Three levels.
When AI helps with several projects, conversation history alone is no longer enough. ORG-STRUCT-23 organises the work in three places: rules, a case registry and project documents. Each answers a different question.
1. SI — how do we work?
The Superinstruction (SI) is a concise starting point. It defines working methods, boundaries, communication rules and links to necessary procedures. It contains rules shared across projects; the details of an individual task belong elsewhere.
2. Registry — what is happening with the case?
Each case has a stable identifier and one owner. The registry holds its current status, the source of decisions, confirmed completion, any blocker and the next step. Before resuming, the AI checks this record to avoid duplicating a task or restarting cancelled work.
3. Project document — where are the details?
The project document describes the objective, source material, parameters, pending corrections, results and verified procedures. It can be a text document or a spreadsheet. One project may cover many cases, so each small correction does not need its own project.
Example: changing a website gallery
- Read the current rules, existing case and project document.
- Check who owns the case, which scope was approved and whether someone is already working on it.
- Prepare the change, keep the necessary backup and perform only the agreed scope.
- Check the published page: images, descriptions, links and mobile layout.
- Record the actual result and evidence. If there is a blocker, state the specific cause and next step.
Why divide the work this way?
A concise SI makes it easier to start, the registry shows the current status, and the project document preserves the knowledge needed to do the work. Stable identifiers let us return to the right item without resending the entire conversation. Fewer copies mean fewer conflicting versions to reconcile.
A record must be read
Saving instructions does not automatically update every open conversation. At the start and after a break, the AI must read the current sources. A plan, a prepared file or a sent assignment is not in itself evidence that the task was completed.