How to make your own agent
Tell Sasha in plain words what job an agent should do. This page shows the prompt for a Project Manager agent and what good output looks like.
Status: planned. Today Agent Creator builds the Finance Agent and the Marketing Agent. Agents that you describe yourself come next. This page shows how it will work.
You do not need a template to get an agent. You describe the job in plain words, in Sasha's chat. Sasha already knows your business, so it fills in the rest: where the facts live, who owns what, and how often you review it.
A good prompt has five parts
- The job. What the agent looks after, for example your projects, your bids or your clients.
- Where to look. Usually "everything Sasha knows": your meetings, documents and connected tools.
- The questions it answers each time it runs.
- What you want to see. A chart, a table, a short email.
- What it may do. For example: "draft, but do not send". These are the agent's hands.
Sasha turns the prompt into the three parts of every agent: its mind (instructions, brief and reading list), its hands (the tools you approve) and its notebook (the record of what it found and did).
Worked example: a Project Manager agent
Create a Project Manager agent for us. It should look across everything Sasha knows: our meetings, documents and connected tools. It should find our projects, and each week tell us which projects are on track, at risk or late, and why. Show each plan as a Gantt chart, marking which dates are confirmed and which are estimated. List overdue actions and their owners, the decisions that are waiting, and new risks. Give me at most three things to do, and draft a chase message for each overdue action. Do not send anything yourself.
What Sasha does next
- It searches what it knows for projects: meeting series, project folders and boards in your task tool. It shows you the list and asks which ones to follow.
- It proposes a plan for each project from what it read. Each date is confirmed (from a task tool, or from you) or estimated (with the sentence it came from). You confirm the plan, and it becomes the baseline.
- It asks a few questions about the gaps, for example who owns a project when no meeting says so.
- It names the agent (a Project Manager agent becomes the Project Agent, unless you choose another name), runs it once for you to check, and then puts it on a weekly schedule.
The Project Agent will also be a ready-made template. See Project Agent.
What good output looks like
These are hand-made examples for Northwind Advisory, the demo business, as on Monday 12 October 2026. They come from Northwind's quarterly review with its client Harbour Foods.
1. The plan, as a Gantt chart
Why this is good practice:
- Every date says how sure it is. Solid bars are confirmed. Dashed bars are estimated, and each one links to the sentence it came from.
- It reports against a baseline. The plan the client agreed is the baseline, so a slip is a fact: "the full report moved from 30 October to 9 November". Measuring progress against an agreed baseline is a core method of PMBOK and PRINCE2.
- Today is on the chart, so what is late is visible at a glance.
2. The weekly email
Projects, week of 12 October: 1 at risk, 1 not complete
Harbour Foods depot network review: at risk. The options work needs the Leeds shift data, which was due on 30 September and has not arrived. The depot workshop is on 21 October, and the list of who attends was due on 7 October. The end date is 9 November, 10 days later than the first plan, as agreed on 18 September at no extra cost.
Castlegate Housing: not complete. The first meeting was on 8 October, but Sasha has no notes of it. Add the notes, and next week's email will include this project.
Do these next
- Ask Siân Price for the Leeds shift and overtime data. A draft message is ready. Tell Sasha: show me the chase message for the Leeds data.
- Ask Priya Nandakumar which depot managers will attend the workshop on 21 October.
- Decide what to do if the Leeds data is not here by 14 October: cost Option A as a range, or move the workshop.
Why this is good practice:
- The answer comes first. The portfolio line says how many projects need you before you read any detail.
- Only the exceptions. On-track work is one line. This is "management by exception", a principle of PRINCE2.
- At most three actions, each with an owner. A long list of actions is a list that nobody acts on.
- A gap is named, not hidden. "Not complete" means that Sasha could not read something, never that all is well.
3. The risk register
| Risk | Likelihood | Impact | Owner | Response | Source |
|---|---|---|---|---|---|
| Leeds labour costs are wrong without the August and September data | High | High | Daniel Okafor | Chase the data; if it is not here by 14 October, show Option A as a range | Meeting, 18 September |
| The Carlisle manager hears about Option B for the first time at the workshop | Medium | High | Priya Nandakumar | Priya and Alan brief him before 21 October | Meeting, 18 September |
| Option C has no site and no lease cost yet | High | Medium | Daniel Okafor | Show a range for the property cost; visit the two sites before the board | Meeting, 18 September |
| Option B routes for the Lake District start at 4 am in winter | Medium | Medium | Daniel Okafor | Test the routes with the Preston manager at the workshop | Meeting, 18 September |
Why this is good practice:
- Each risk has an owner and a response, not only a worry. Risk standards such as ISO 31000 expect each risk to have an owner and a chosen treatment.
- Each risk has a source, so you can check why it is on the list.
- Likelihood and impact are separate, so a likely but small risk does not hide an unlikely but serious one.
4. A chase message
Hi Siân, I hope the stock count went well. At our review on 18 September you kindly offered to send the Leeds shift patterns and overtime hours, by week, for August and September, once the September payroll had run. Could you send them by Wednesday 14 October? They let us cost Option A fairly before the depot workshop on 21 October. If a part of the data is hard to get, the overtime hours alone would still help. Thank you, Daniel
Why this is good practice:
- It is suggested text. The agent never sends it. A person reads it, changes it and sends it.
- It reminds the person what they agreed, and when. It does not blame.
- It asks for one thing by one date, and says why it matters.
- It offers a smaller fallback, so the work can go on.
Two more prompts
Create a Bid Manager agent. Each week, list the tenders we are working on, their deadlines, what is still missing from each bid and who owns it. Warn me ten days before a deadline when something is still missing.
Create a Customer Success agent. Each month, list our clients whose contract renews in the next 90 days, and say how each relationship looks from our meetings and emails. Name the clients who may be at risk, and suggest one next step for each.
Every agent learns from a reading list
Each agent comes with a short reading list: the methods it uses and where they come from. While it runs, the agent suggests new sources and methods, and you choose which to add. See How agents learn.