The Founder’s Personal AI Agent Kit
Stop briefing your AI from scratch. Give it context, boundaries, and two jobs worth remembering.
Start with a small, editable record of your business and two recurring jobs. Choose a tool that can retain that context, access the sources those jobs need, and run at the required times. Verify each ability separately. Keep messages and consequential actions behind your approval, review what it remembers, and keep an export you control. A promise to remember is not proof that a watch is running.
Read time: approximately 10 min
Why are you still explaining your company to an assistant you used yesterday?
You ask for an investor follow-up. Before it can help, you explain the company, the round, what the investor asked last time, and how you write. Tomorrow, a client email needs a reply. You explain a different slice of the same business again.
The writing is faster. The briefing is still your job.
A useful personal-agent setup reduces that repeated briefing and gives a few important jobs somewhere to live. It knows which context to retrieve, which sources it can inspect, and when to bring something back to you. None of that requires handing over your inbox on day one.
This playbook is for founders who already use AI for answers and drafts and want a more reliable assistant. It is a setup framework, not an installation guide for a particular product. You need an approved tool, access to its memory and scheduling settings, and a way to check what it actually did.
1. Choose the setup by the job, not the label
“Chatbot,” “DIY agent,” and “personal agent” are useful starting labels, but they are not clean capability boundaries. Do not assume a chat product forgets everything, or that anything sold as an agent watches your business continuously.
| Setup | Good starting use | What to verify |
|---|---|---|
| Chat assistant | Answers, drafting, and work you initiate. | Whether saved context persists, what it can access, and whether scheduled tasks are supported. |
| DIY agent setup | Workflows where you want control over files, rules, integrations, and hosting. | Who maintains it, how jobs run, and which data survives a restart. |
| Managed personal agent | A lower-maintenance assistant with supported memory and connected workflows. | Actual integrations, background execution, approval controls, export options, and plan limits. |
| Self-hosted or local setup | Work requiring more control over where components run. | Whether model requests or connected services still send data elsewhere. Local software alone does not mean local processing. |
If you only need better drafts, improve the context you give your existing tool first. If your problem is missed follow-ups, test its scheduling and source access. If you need a particular folder structure or stronger control over storage, consider a DIY setup. More maintenance is a cost, not a badge of seriousness.
2. Before you connect anything
Choose one job you already do manually. Investor-call preparation or a reminder about unanswered deal emails is enough. Write down what a correct result would contain and where the underlying information lives.
Then check four things: can the tool read the required source, retain the necessary context, run without an open conversation, and deliver the result somewhere you will see it? A missing capability means a different setup or a narrower job, not a more forceful prompt.
Start read-only where possible. Drafting a reminder and sending a reminder are different permissions. You can gain most of the first week’s value without allowing external messages, purchases, cancellations, or changes to business records.
3. The founder memory file
Keep a master copy named founder-memory.md, or an equivalent editable document you control. The filename is a convention, not a universal product requirement. Some tools use files; others store memory internally. Confirm how your chosen tool loads and updates it.
# About me
- Name and role:
- Company, in one line:
- Stage and team size:
- Time zone and working hours:
# This quarter
1.
2.
3.
# People and relationships
- Investors: name, firm, current questions
- Key clients: account, what they buy, renewal date
- Co-founders and key hires: name, responsibility
- Store only information I am permitted to share.
# How I work
- Protected time and meeting preferences:
- Approved sources: email, calendar, notes, CRM
- Decisions to flag rather than make:
- Actions that always require my approval:
# My voice
- Email tone:
- Phrases I use:
- Phrases I avoid:
# Reply rules
- Lead with the answer, then the detail.
- If I have actions remaining, end with “What’s next.”
- List only actions I actually need to take.
- Say when a source is unavailable or a fact is uncertain.
# Maintenance
- Last reviewed:
- Ask before changing sensitive or consequential facts.
- Keep temporary task status out of long-term preferences.
- Never store passwords, API keys, or access tokens here.The rest of this playbook is below.
Enter your email to read it. You also get one new playbook every week.
One email a week. Unsubscribe any time.
The most useful lines are often the boundaries. “Flag decisions rather than make them” is more actionable than “be my chief of staff.” So is naming the sources it may read and the actions it must not take without approval.
Do not turn the file into an archive of every conversation. Keep stable context here. Put dated work and changing deadlines in the relevant task or project record, with a source the assistant can check.
4. Three prompts for the first session
The interview. Use this when filling in the template feels like more work than answering questions.
Interview me to build my founder context. Ask one question at a time about my company, priorities, key relationships, working preferences, and recurring tasks I forget. Do not ask for credentials or information I am not allowed to share.
After ten questions, show me a draft memory summary. Separate confirmed facts from assumptions. Ask me to approve it before saving. If you cannot persist memory, give me an editable copy instead.The rule-setter. A correction is useful only if you can verify where it went.
My preference is: [rule]. Show me the exact wording you would save and its scope. After I approve, save it in the supported persistent context and identify where it was stored. Tell me if it conflicts with an existing rule. Do not claim it will apply in every future session unless the system actually supports that.The weekly reset. Review changes before they become permanent context.
Review this week’s changes using only the calendar, email, and notes I have explicitly connected and authorized. Propose updates to my priorities, relationships, and deadlines, with a source for each material change. Separate temporary task status from lasting preferences. Show the proposed changes for approval before updating memory, and list sources you could not access.5. Start with two watches, not seven
A watch is a defined check with a trigger, source, and destination. “Keep an eye on my business” is not one. The examples below are specifications to adapt, not jobs that activate merely because you paste them into a conversation.
| Watch | Copyable request | Required source |
|---|---|---|
| Overdue invoices | Every weekday at [time, time zone], flag unpaid client invoices more than [7] days overdue. Include the invoice link and draft a polite reminder for approval. Do not send it. | Authorized billing records with due dates and payment status. |
| Investor prep | At [time, time zone] the evening before an investor call, prepare a one-page brief with sourced metrics since our last call and unresolved questions. Mark unavailable information. | Calendar, approved metrics, and prior meeting notes. |
| Client emails | During [working hours, time zone], check for new emails from [clients] every [interval]. Alert me once per new thread with a two-line summary and link. | Authorized email access and a supported checking cadence. |
| Software renewals | Check our subscription register daily. Warn me [7] days before renewal with the recorded price. Include usage only if an authorized source provides it. Do not cancel anything. | Subscription register; usage data if available. |
| Deal follow-ups | Every weekday, find deal threads I sent that have no reply after [5] days. Check the full thread, remind me once, and draft a short follow-up. Do not send it. | Sent and received messages plus a way to identify deal threads. |
| Hiring | At [cadence], flag new replies from candidates for [role]. Link to each reply and propose an update to my hiring notes for approval. | Authorized candidate correspondence and hiring notes. |
| Price drops | Check [product URL] at [cadence]. Alert me when the total price in [currency], including known fees, falls below [amount]. Never purchase it. | Accessible pricing page; availability and variant information. |
For every watch, require a setup receipt: the trigger, time zone, connected sources, delivery destination, approval boundary, and how to pause it. Set an alert limit and a duplicate rule. Missing data should produce a visible failure notice, not a confident “nothing to report.”
6. The part that breaks: remembering is not running
An assistant can correctly remember that you want invoice reminders and still never check an invoice. Memory is context. A scheduler or supported event trigger starts the work. Source access supplies the evidence. Delivery gets the result to you.
These are separate pieces. Test them separately.
Test memory. Start a fresh conversation and ask it to retrieve a harmless preference you saved. Check that the answer comes from stored context rather than information you repeated in the question.
Test execution. Use an authorized test item and a near-term trigger. End the conversation. Confirm that the job ran without another message from you. If it is meant to work with your laptop closed, test that condition too.
Test evidence. Check one output against the underlying source. A fluent investor brief that cannot access your metrics is not a successful run.
Test delivery and failure. Confirm where the result arrived and what happened when a source was unavailable. A blank report must not masquerade as a completed check.
Keep a small execution record separate from your memory file: last successful run, sources checked, failures, and alerts already delivered. This helps distinguish a quiet business from a broken connection.
7. Privacy and portability before convenience
Read the selected product’s current data policy and your plan’s settings. Check retention, training use, connected-app permissions, deletion, and export. Do not assume the same rules apply to every plan or integration.
- Connect only the accounts required for the first two jobs.
- Check client contracts, NDAs, and company policy before sharing business data.
- Keep passwords, banking credentials, API keys, and tokens out of memory and prompts.
- Require approval before external communications or consequential changes.
- Keep an editable master copy of your founder context and export supported memory regularly.
- Test that the export is readable and useful elsewhere. Plain text may transfer context, but it does not automatically transfer integrations, schedules, or permissions.
- For sensitive or regulated work, involve the appropriate security, privacy, or legal owner before connecting data.
A local agent is not automatically private. It may still call a hosted model or external services. Model capability also depends on the model and hardware you choose, not simply whether the agent runs on your machine.
8. Your first week
| When | Do this | Evidence it worked |
|---|---|---|
| Day 1 | Choose one setup against two real jobs. Review its data policy, permissions, and limits. | You know what it can read, remember, run, and deliver. |
| Day 2 | Fill the memory template and run the interview. Approve the summary. | A fresh conversation retrieves a harmless saved preference. |
| Day 3 | Connect only the sources the selected jobs need. Run each check manually. | The output links to the right records and names missing data. |
| Day 4 | Configure two watches and perform an unattended test. | A run record and delivered result exist without another prompt. |
| Day 5 | Correct mistakes, remove noisy alerts, and verify the approval boundary. | No unwanted messages or business changes were made. |
| Weekend | Run the weekly reset, review proposed changes, and export the context. | The export is readable and you can explain what changed. |
After two weeks, decide whether the assistant reduced repeated briefing or caught something worth acting on. Keep the jobs that helped. Pause the ones that created another inbox.
9. If something goes wrong
It forgets your preferences. Was the context actually saved, and does this session load that store? Ask for the stored entry rather than another promise.
A watch never fires. Does a real scheduled job or event trigger exist? Check its enabled state, time zone, runtime, and last-run record.
It sends repeated alerts. Does the job retain which items were already reported? Add an identifier-based duplicate rule and a delivery limit.
The brief looks complete but misses information. Check source permissions and freshness. Require an explicit list of unavailable sources.
Old priorities keep appearing. Review the master context for conflicting or outdated entries. Approve a correction instead of adding another contradictory rule.
10. What this changes—and what it does not
A personal agent does not remove the need for judgment. It can prepare a follow-up, but you still decide whether the relationship needs one. It can surface a renewal, but it cannot infer whether the team will need that tool next quarter from a price alone.
The useful shift is smaller: less repeated briefing, fewer forgotten checks, and clearer evidence when something needs your attention.
Start with a memory file you can inspect, two jobs you can test, and boundaries you can explain. If those work, expand. If they do not, adding five more integrations will only make the failure harder to understand.
