Organise an AI project around one repeated result, not around everything the model might know. Give it a purpose, a small set of approved sources, a precedence rule, a fixed output brief and a named human review point. Add connections only when a live task needs them.
The smallest useful structure
Create these seven records. They can begin as short files; completeness is less important than knowing which one owns each decision.
| Record | What it controls | Done when |
|---|---|---|
| Project canvas | Audience, job, owner and finish condition | One recurring result is named |
| Source register | Approved facts, dates and limits | Material claims point to a current source |
| Offer or subject canon | Current terms, exclusions and links | Old drafts cannot silently override it |
| Project instructions | Precedence, refusal and review behaviour | Missing evidence produces a question, not invention |
| Asset brief | Reader, output, evidence and acceptance checks | Another session can reproduce the required shape |
| Connection card | Data scope, permissions and revocation | Denial and removal fail safely |
| Review record | Corrections, outcome and next decision | Continue, reshape, pause or stop is recorded |
A ten-minute first installation
1. Name one result you expect more than once, such as a Monday brief or a source-grounded page outline. 2. Add only the files needed for that result. 3. Write which file wins when two records conflict. 4. State what the project must never send, publish, purchase or change. 5. Run one normal case and one case with a missing or stale source. 6. Keep the project only if both outputs are inspectable and the failure case is honest.
Do not connect a calendar, drive or mailbox simply because the option exists. First prove the output manually with a bounded input. A connection earns its place when it removes repeated retrieval work without expanding the project's authority beyond the result.
What repeated result should this project produce?
Keeps the project tied to one recurring job rather than a miscellaneous knowledge dump.
Which files are allowed to supply facts and which record wins when they conflict?
Establishes source precedence and exposes missing authority.
What happens if a plausible output is wrong or stale?
Sets the review and permission ceiling.
Can a named person inspect the result before it changes external state?
Separates a useful project from unsafe delegation.
When a project is the wrong layer
A one-off task normally needs a better prompt, not a maintained project. A stable sequence spanning several systems may need a controlled workflow. Variable tool choice and external action may need an agent assessment. If no reliable review point exists, keep the task human-led.
Frequently asked questions
Use one recurring job, a small approved source set, explicit precedence, a fixed output contract and a human review point. Keep temporary or sensitive material outside unless the task genuinely needs it.
No. A disconnected project is often safer and simpler. Add one read connection only after the manual version proves useful, then test denial, expiry and revocation.
Only enough to make the current result reproducible and challengeable. Remove superseded files instead of allowing several versions to compete.
Clarify the right AI layer
The short interview distinguishes a prompt fix, maintained workflow, bounded agent and human-led stop before recommending a product.