The failure mode of AI tooling at work is rarely the model. It is that the context lives in six places: a chat thread here, a PDF there, a spreadsheet, a to-do list, a bookmark folder and someone's memory. The work becomes re-explaining the project every morning.
Project-scoped, not chat-scoped
A chat thread is a bad container for work because its boundary is time, not topic. A project is the right container: one place for the sources, the working notes, the open questions and the outputs.
When you start, create the project first and put everything that defines it inside:
- The brief, as a document - not as a message in the thread.
- The source files, uploaded once with names that mean something.
- The constraints: budget, deadline, audience, things that are out of scope.
Then the conversation happens inside that boundary, and every answer inherits it.
Make the model read before it writes
Most disappointing outputs come from a model that was asked to write without being given the material. Two habits fix it:
- Summarise the sources first and check the summary. If the summary is wrong, the draft will be too.
- Ask for the argument before the prose: "list the five claims you would make, with the source for each".
The second habit turns a generation task into a review task, which is where your time is actually worth spending.
Files in, documents out
Uploads are only useful if the output can cite them. Ask for the claim and the file it came from, and keep a section for "not supported by the sources" so the gaps are visible. When the deliverable is a document, generate it as a file you can download and edit - a chat message is not a deliverable.
Scheduled work beats remembered work
Recurring tasks - a weekly digest, a competitor check, a status summary - should be scheduled in the tool rather than remembered by a person. The value is not the automation itself; it is that the task stops depending on someone having a free Friday.
Keep the audit trail
Keep the prompts and the outputs together in the project. Six weeks later, the question is never "what did it say" but "why did we decide that", and the answer is in the thread you did not delete.
Tooling
This is what a workspace with conversations, projects and scheduled tasks in one place is for - for example Menus AI. The tool matters less than the boundary: one project, its sources and its outputs in the same container.
The rule
Open the project, put the sources in it, summarise before writing, keep the trail. Re-explaining the project every morning is the cost of not doing it.
Top comments (0)