Getting better results

You know the basics now. You can talk to your assistant, connect your tools, and set up workflows. This chapter is about getting genuinely good at it.

The difference between a workflow that kind of works and one that nails it every time usually comes down to two things: how much your assistant knows about you, and how clearly you write the instructions. The first one builds over time. The second one is a skill you can sharpen right now.

Specific beats vague

We touched on this in Chapter 1, but it is worth going deeper because it is the single most important skill.

Compare these two instructions for a morning briefing workflow:

Summarize my emails.

Your assistant has to guess at everything. How many emails? How far back? What format? What counts as important? You will get something back, but it probably will not be what you wanted.

Now compare:

For each unread email from the last 12 hours: give me a one-line summary, the sender’s name, and the time received. Flag anything that needs a reply as ACTION NEEDED. Group the rest under FYI. Skip newsletters, marketing emails, and automated notifications.

Same task. But the second version tells your assistant exactly what to produce, what to skip, and how to organize it.

This is the same principle as delegating to a colleague. “Can you put together a report?” is going to get you something very different from “Can you put together a one-page summary of Q1 revenue by region, with a comparison to last quarter, by end of day?”

Your assistant knows a lot about you already - your role, your contacts, your priorities. But it cannot read your mind about output format. Tell it what you want.

Tell it what to skip

Your assistant does not know your preferences unless you state them.

If your inbox gets a lot of newsletters, say so: “Skip anything from a mailing list.” If you do not care about calendar events that are just focus time blocks, say so: “Ignore events I created for myself.”

Defining what to exclude is often more valuable than defining what to include. Your assistant can figure out what is important - it knows your work. What it cannot figure out on its own is what you do not care about.

Explain why, not just what

When your assistant understands the purpose behind a task, it handles ambiguous situations better.

Say you have a workflow that flags urgent emails.

Flag urgent emails.

What counts as urgent? Your assistant has to decide, and it might decide differently than you would.

This workflow helps me make sure nothing time-sensitive falls through the cracks. Flag emails that have a deadline within 48 hours, require my input before someone else can proceed, or are from my CEO. When in doubt, err on the side of flagging - I would rather review a false positive than miss something real.

Now your assistant has a decision framework. It knows what “urgent” means to you, and it knows how to handle edge cases. The more context you give it about your reasoning, the better its judgment becomes.

Drafts that sound like you

One of the first things people notice about Town is that drafts actually sound like them. This is not magic - it is your memory profile at work. Your assistant learned your voice from your emails during onboarding, and it keeps learning from every correction.

But you can make it even better. Here is a before and after.

Say a vendor emails you about renewing a contract, and your assistant drafts a reply:

Dear Sarah, Thank you for reaching out regarding the contract renewal. I have reviewed the proposal and would like to discuss some modifications to the terms. Would you be available for a call this week to discuss further? Best regards.

That is fine. It is polite, it is professional. But it does not sound like you.

Now say you have been correcting your assistant’s drafts for a week. You have told it things like “I do not say ‘Dear,’ I say ‘Hi.’ I keep things short. I propose next steps, I do not just ask for a call.” The next draft looks different:

Hi Sarah, Thanks for sending the renewal over. Reviewed it with the team - we are aligned on scope but want to adjust the timeline. Can we push the start to Q3? Happy to discuss this week.

My calendar’s here: [link]

Same information. But it sounds like a person - specifically, like you. This is what “gets sharper over time” feels like in practice. Every correction teaches your assistant something about your voice, and it compounds. The drafts get noticeably better, week over week.

Four workflow shapes

Most workflows follow one of four patterns. Recognizing which one you are building helps you make better decisions about triggers, tools, and trust levels.

The classifier. It watches a stream of incoming items and sorts them. Every email gets a label. Every receipt gets logged. It processes fast and moves on.

Auto-label is a classifier. It runs on every email, makes a quick decision, and takes a lightweight action. Classifiers should almost always be autonomous - they are high-frequency and low-risk.

The reporter. It gathers information on a schedule and delivers a summary. Your morning briefing is a reporter. So is a weekly pipeline report or a daily standup digest.

The key to a good reporter is being specific about what to include, what to skip, and how to format it. Because your assistant knows your priorities, a well-written reporter feels like it was made just for you - because it was.

The watcher. It monitors for a specific condition and fires when it matches. “When a meeting starts in 15 minutes, send me a prep doc.” “When someone emails me from a domain not in my CRM, create a contact.”

Watchers are more targeted than classifiers. They do not act on everything - they wait for the right signal.

The responder. It handles incoming requests end to end. Someone asks to schedule a meeting, and the workflow checks your calendar, finds available times, and drafts a reply in your voice. Someone sends an invoice, and the workflow extracts the details, logs them, and drafts a forwarding email.

Responders do the most work and benefit the most from approval-required mode. Let them handle the research and drafting, then you review the final output.

Combining patterns

Real workflows often combine two patterns.

An invoice logger is a classifier (is this an invoice?) that becomes a responder (extract the data, log it, draft a forward). A meeting prep workflow is a watcher (event starting soon) that becomes a reporter (here is what you need to know about the attendees).

Once you recognize the patterns, designing new workflows gets much faster. You know which trigger to pick, which tools to enable, and where to set the trust dial.

Progressive trust

This is what experienced users do, and it is worth stating explicitly.

When you create a new workflow, start in read-only. Run it for a day. Look at what it would have done in the session history.

If it looks right, promote to approval-required. Run it for a week. Approve or reject each action. Correct the instructions when something is off.

If it has been reliable for a while, promote to autonomous. It just runs.

You can always dial back. This is the same arc as the delegation analogy - small tasks first, then bigger ones, with checkpoints along the way.

Power moves

A few things experienced users do:

• Memories persist. Tell your assistant a preference once - “I always want meeting prep 30 minutes before, not 15” - and it remembers across all future interactions. You do not need to repeat yourself. These preferences stack up over time and are a big part of why your assistant keeps getting better.
• Workflows can call other workflows. Build small, focused pieces and chain them. A research step, a drafting step, an approval step - each one does one thing well.
• The sandbox runs code. Your assistant can write and execute Python for data processing, calculations, file manipulation, and chart generation. If a task involves crunching numbers, it can handle it.
• Session history shows everything. Every workflow run is logged in full detail - what it read, what it decided, what it did. This is your audit trail and your debugging tool.

What’s next

You now know how to write better instructions and recognize common patterns. The last chapter covers how to fine-tune your assistant’s behavior, fix what is off, and keep everything running the way you want it.