A proposal is ready. The pricing sheet is finished. Just before the files go out, someone adds a condition: only one location will join the first phase; the other two will be considered after the trial.
Changing the number in the proposal is easy. The pricing sheet, however, still charges for three locations. The rollout plan still schedules three training sessions. An old version remains among the attachments. Each document looks plausible on its own. Together, they describe three different engagements.
Work like this rarely starts with a blank page. It starts with a folder: an earlier proposal, a few business spreadsheets, an export of recent conversations, and several attachments with “final” in their names. Much of the effort goes into understanding how these materials relate, then carrying a change into every place it affects.
Cowork’s research preview brought this kind of work to a more accessible desktop entry point in January 2026. The official release notes describe a January 12 preview for Max subscribers on macOS, running in an isolated virtual machine with direct access to local files and MCP integrations. Access expanded to Pro subscribers on January 16. That gives us a concrete starting point for a broader question: how should we hand over a complete task when AI can work with files?
A request needs a working context
Ask an AI assistant to draft an implementation proposal, and it can quickly produce objectives, phases, staffing, and a schedule. A complete-looking document may still fail to connect with the work already underway.
Which pricing version applies? What support was promised before? Is this an extension of an earlier project, or a fresh discussion of scope? The answers tend to be scattered across the materials. Uploading every file does not explain which documents provide background and which still govern the work.
The person assigning the task often knows these distinctions without noticing the judgments involved. “Refer to the last proposal” may mean “borrow its structure.” A model could also carry over its service scope. Adding more files makes that ambiguity harder to resolve with a request to read carefully.
A more useful preparation is to explain each document’s role in this task. For example: reuse the earlier proposal’s section structure; take the deliverables from the current scope note; calculate fees using the applicable price list. The earlier total and support period provide historical context only.
A short task brief can give an ordinary folder enough context for work to begin. There is no need to design a perfect knowledge base first.
A folder can have two maps
The first map is familiar: proposals go in one folder, prices in another, and meeting records are arranged by date. It answers where things are stored.
A task also needs a second map: which conditions determine which parts of the work?
In this example, “one location in the trial” affects the proposal’s scope, the quantity used in pricing, and the training arrangements in the rollout plan. Those details live in separate documents but share a premise. Generating each document independently makes it easy for that premise to drift.

This map need not become a complex system. A few sentences in the brief, or a column identifying the basis for each part of the output, can serve the same purpose. Making the relationships explicit helps the person doing the work and the person receiving it understand the same task.
It also helps limit what needs to be read. To revise a one-location trial proposal, begin with the current scope, applicable prices, and the relevant earlier proposal. Unrelated market research and historical projects can stay outside the task. A small set of materials with clear roles is easier to inspect than a large collection with ambiguous boundaries.
The map itself can be wrong. A fee omitted from the scope note does not become free because there is no arrow for it. The purpose of the map is to expose relationships and gaps so they can be checked against the business materials.
The revealing part is the second revision
Producing a first set of files is easy to demonstrate. What happens after a condition changes is more revealing.
Suppose the responsible person confirms that the engagement will expand to three locations after the trial. The price list and charging rules remain unchanged. This requires more than replacing “one” with “three” throughout the files. The proposal needs the expanded scope. Pricing must be recalculated under the applicable rules rather than mechanically tripled. The rollout plan needs fresh decisions about training, staffing, and dates.
Meanwhile, the earlier proposal should remain intact as a historical reference. The price list also stays unchanged: changing a quantity does not change the pricing rule.
To make the distinction clear, this example keeps scope in the proposal, fees in the pricing sheet, and specific dates in the rollout plan. If a real proposal also contains dates or amounts, those occurrences belong in the revision scope too, and the matrix below must change accordingly.

A more useful instruction is to update the affected content from the confirmed change and identify the consequences across documents, with the confirmation supplied as its basis. That defines a bounded revision task.
If the AI cannot determine whether training must happen in three separate sessions, it should leave that question visible. A scheduling judgment is missing. Regenerating a polished proposal will not supply the answer.
After the revision, read across the documents using the same business condition. If the scope covers three locations, does the fee correspond to three? Does the training plan cover them as well? This is how we can find a set of documents that reads smoothly one by one but fails as a whole.
It also gives us a practical way to assess whether AI has helped. After a change, how many old statements still need to be found manually? Can the next person understand what changed and why? Those questions are closer to the actual burden of the work than the number of pages generated.
The next task should need less explanation
After a revision, the final proposal is the easiest thing to save. Yet the next task may need something else.
Why did one training arrangement fail to fit? Which fee needed a separate assessment? How was an ambiguous sentence rewritten? If those lessons remain only in the conversation, they have to be explained again next time.
One useful distinction is between facts belonging to the current engagement and methods that can travel to another task. The location count and confirmed dates belong in project materials. A requirement to explain what a quote excludes, or to reassess training when an engagement expands, may be useful again.
Reviewed examples help too, especially when saved with a reason for using them. “This page helps a location manager understand the trial scope quickly” is more informative than “This is an excellent proposal; copy it.” The first describes a purpose. The second invites old facts to travel with the format.
We want AI to draw on such materials to understand how a team works: which documents provide lasting guidance, which situations require a new judgment, and what makes a result usable for the next person. Ordinary files can support that effort without requiring a complete change of tools.
What gets retained still needs care. A discount granted for one project should not become a general pricing habit. A temporary preference should not automatically become a team rule. Each lesson needs a sense of where it applies.
File tasks are appealing because AI can participate in connecting materials, revising them, and carrying work forward. That may reduce copying and reassembly. It may also spread a misunderstanding faster. The difference often lies in relationships that previously existed only in someone’s head.
For a first attempt, choose work with existing materials that will continue to change. Begin with three or four related files, explain their purpose and authority, produce a version, then take it through a real revision. That makes it easier to see which work can be delegated, which business judgments still need to be supplied, and which lessons are worth keeping for the next task.


