Notes: Stop explaining the same work from scratch

The situation. Two days after the Northstar client review, Alex has a new check-in transcript and another round of progress updates to write. The 3-sentence preference from the first section is still sitting in Friday's conversation.

Searching old threads for the prompt that worked is one way back to it. When Alex finds the thread, though, last meeting's dates and finished items are mixed in with the two instructions worth keeping.

So I sort what Alex needs into two piles:

• details that change every week: the new transcript, where the role list stands, the next check-in date
• parts that should survive: the Northstar background and the follow-up method

With those piles kept apart, the next meeting starts like this:

1. Alex supplies only the new transcript and the goal for the upcoming conversation.
2. The agent opens a saved file of Northstar facts, such as who Priya and Jordan are and where the role list stands.
3. It drafts with Alex's saved follow-up method, and Alex retypes no instructions.

Alex's attention goes to what changed this week, instead of rebuilding what already worked last week.

The 3 scopes. The company folder holds background that applies everywhere, such as what your organization delivers. The department folder holds the context for Alex's Delivery Operations team. Inside it sits a project folder for one piece of work, Northstar Onboarding.

You do not write those files from scratch. Alex supplies permitted notes, or explains the setup aloud, and asks the agent to draft them.

Inside the Northstar project the agent keeps 3 records, each with its own job:

1. Project instructions: how to approach work here, such as checking delivery dependencies before drafting client updates.
2. Current facts: today's situation, such as the role list status, and the team: Priya leads delivery, Sam coordinates onboarding.
3. Decision record: the October 5 choice to start Tuesday's setup for the 18 confirmed client staff and hold the 4 specialist accounts until Thursday, so nobody guesses security permissions.

Alex's follow-up method does not go in the project instructions. It belongs in a saved skill that works for every client.

Show me what you found. A neat folder tree does not give the agent memory of everything in it. So Alex asks the agent to retrieve the Northstar records and list them before continuing, and checks the list in seconds.

Update in place. When Jordan confirms 18 of the 22 client roles, Alex asks the agent to update the current facts record and show the change. The next Northstar task starts from this week's facts.

Start where it worked. Begin inside the conversation that produced the useful result, while your goal, the agent's questions and your corrections are still on screen. Copying only the final draft loses the reasoning that made it useful.

You do not write a template with blanks to fill in every week. You paste a complete capture request and let ChatGPT or Gemini write the detailed prompt.

The capture request asks for a thorough, versatile prompt that keeps:

• the purpose of the work and the reasoning behind it
• the kinds of input it needs
• the quality criteria that made the result useful
• the rules for adapting when the next meeting or audience differs

In the Northstar illustration, the prompt keeps the rule to check the transcript against Alex's purpose for the next review, and to stop if that purpose is missing. Friday's October 2 transcript and Monday's October 5 review date stay out.

There are 2 routes from a checked prompt to a saved skill:

• From ChatGPT or Gemini: paste the prompt into a ChatGPT Work task with a permitted transcript, check the result, then ask @skill-creator to package it.
• Already in Work: call @skill-creator in the same thread and ask it to capture the method from the conversation.

Here it saves the meeting-follow-up skill, which opens from Skills in the sidebar of the ChatGPT desktop app. If your company does not give you Work, the saved prompt still works in an ordinary chat.

Run it on a new meeting. In Work, Alex types @ to pick meeting-follow-up and supplies the transcript from Wednesday's technical check-in. The skill brings the method, and the transcript and project record bring this week's facts.

Much of the result works first time. The drafts keep the clear action and reason Alex wanted.

The gap. Priya and Sam both have preparation tasks for the security check before Thursday's 2 p.m. session. The draft writes to them together and never says which one confirms readiness back to Alex.

Alex has 3 ways to deal with it:

• fix that line by hand every week, which keeps the supervision going
• paste Thursday's details into the skill, which leaves next week with outdated facts
• add one general rule to the skill and keep the details in the project record

Alex takes the third. The agent proposes one line: when an action involves more than one colleague, state who is assisting and name the single owner who sends confirmation.

A quick rerun on Wednesday's transcript names Sam as the confirmation owner. The skill now saves a correction Alex would otherwise make every week.

The decision. Maya directs proposals for a services firm. Meridian Manufacturing wants its old manufacturing records moved into a new platform, plus a custom two-way connection to its main business system, in 6 weeks. Maya must decide whether to bid on the full scope, propose a narrower phased scope, or decline.

Asking AI to act as a senior commercial strategist and say whether to bid brings back a confident essay that knows nothing about her team's real capacity.

So Maya starts from the option-comparison skill in the starter pack, asks for a commercial and delivery reviewer's view, and defines good work through 4 criteria:

1. delivery requirements
2. specialist capacity impact
3. viable alternatives
4. assumptions that could change the decision

She asks the agent to explain how it will evaluate the proposal and to list missing inputs before it runs the full comparison.

The flaw she catches. The plan compares the 6-week timeline with the total number of engineers on her team sheet. The custom connection needs a senior integration specialist, and both of hers are committed to another client's launch in weeks 3 and 4.

She corrects the plan before it runs: check specialist availability in weeks 3 and 4, and add a phased option that moves integration to week 5. The comparison then uses the right constraint, and Maya can keep that check in her adapted skill for future proposals.

When Alex asks the agent to run Meetings and follow-through on a transcript, the workflow follows its order:

1. retrieve the Northstar project instructions, current facts and decision records, so the agent knows who Priya, Sam and Jordan are
2. read the transcript and pull out promised actions, deadlines and open items
3. leader-interview checks that the purpose of the next review is clear, and asks Alex if it is not
4. meeting-follow-up writes the drafts from the transcript notes, the project background and the confirmed purpose
5. the notes, drafts and reminder proposals go onto one review page for Alex

The common mistake. Listing skill names in a chain and assuming each step knows facts it was never passed. If meeting-follow-up never received leader-interview's output, Jordan's draft would drop the blocker-and-decision question Alex needed for Monday's call.

A well-built workflow also says what to do when something is missing:

• no stated purpose for the next conversation: it pauses at leader-interview to ask Alex
• no calendar access or approval: the Monday 10 a.m. reminder goes on the review page as Proposed only, with a note of what was not done

Lighter maintenance. The workflow reads the same saved meeting-follow-up skill, so the confirmation-owner rule from 2.4 is designed to reach it, as long as both are saved in the same place.

This section named the pieces that keep AI work from starting at zero:

• a project keeps the Northstar background and current facts in one place
• a decision record keeps each confirmed choice with its reasons
• a saved skill holds the follow-up method, and one targeted line improved it
• an expert method plus a checked plan got Maya through an unfamiliar decision
• skill orchestration lets one request run several skills in order

When the next meeting arrives, the new transcript and the updated project record supply this week's facts, and the saved skill supplies the method that already proved useful.

Because the skill keeps the reason behind each instruction, such as why a recipient needs to know the purpose of a request, the agent can adapt when the topic moves from an access list to a budget check.

The new pressure. Summaries, drafts and updates now arrive much faster, for you and your colleagues. Your inbox fills with choices competing for your attention, and the next section is about protecting it.