Build Your AI Co-Founder Agent: The Founder's Playbook

Everyone is racing to build ten AI agents. One for ads. One for email. One for content. One for search. Then they wonder why nothing improves.

The agents work in isolation. Nobody connects their output. You end up with five reports landing in your inbox and zero clarity about what to do today.

The fix is to build one agent first: an AI co-founder that sits above the rest. It learns how you think and what your business actually needs. It runs your specialist agents. It reads everything they produce and hands you the two things worth acting on. It remembers the calls you made last month, and when you're about to do something dumb, it says so.

This is the playbook to stand one up today. By the end you'll have the exact system prompt, two context files, the sub-agent wiring, a decision log, and a weekly cadence — copy-paste ready. Build this agent first, and every other agent you add gets easier, because the brain at the center already understands the business they work for.

Why the co-founder agent comes first

Adding more agents does not automatically make you smarter. The honest verdict from teams running multi-agent systems in production is that most "more agents = more intelligence" gains were an illusion — the agents just rearranged the same information. When these systems fail, they fail for a structural reason: no single agent owns the big picture or the final call.

The pattern that actually survives production is called the supervisor pattern: one orchestrator agent that holds context, delegates to specialists, and synthesizes their work into a decision. That is exactly what an AI co-founder is. It is not a fleet of equals talking to each other. It is one brain at the center with specialists reporting up.

Agents are not a fad you can skip, either. Analysts expect roughly 40% of enterprise applications to embed AI agents by the end of 2026. The founders who win won't be the ones with the most agents. They'll be the ones whose agents are coordinated.

So build the coordinator first.

How the co-founder agent is built

Five parts. Each one is a copy-paste asset below.

• The system prompt — defines the agent as an operating partner, not a chatbot. Its job is to synthesize, prioritize, and challenge you.
• Layer 1 — "It knows you" — a context file on how you actually operate, so it advises through your lens.
• Layer 2 — "It knows your business" — a context file on your model, numbers, and team, so it audits like a senior operator every week.
• The orchestration spec — the map of your four specialist sub-agents and the prompt that collapses their reports into a decision.
• The push-back block + decision memory — instructions that make it disagree with you when you're wrong, and a log so it never forgets a call you made.

Then a weekly cadence ties it together, and build steps stand it up. Let's go.

Part 1 — The system prompt

This is the core. Paste it into your agent's instructions exactly. It tells the agent it is an operating partner, gives it its mandate to synthesize and prioritize, and orders it to challenge you.

promptCopy
# ROLE
You are my AI co-founder and operating partner. You are not a chatbot, and not
an assistant that waits for orders. You act like a co-founder who owns outcomes
with me. You think in terms of leverage, focus, and what moves the business.

# CONTEXT YOU ALWAYS USE
Before you answer anything strategic, read and reason through these files:
- founder.md   — how I operate: my strengths, limits, time, and goals.
- business.md  — the model, offers, numbers, team, and this quarter's focus.
- decisions.md — every major decision, its outcome, and the lesson learned.
Treat these as the source of truth. If my request conflicts with them, say so
before you do anything else.

# YOUR MANDATE
1. SYNTHESIZE. You read the reports from my specialist agents (ads, search,
   email, content) plus the numbers, and you collapse them into the few things
   that actually matter. I never want ten dashboards. I want one clear answer.
2. PRIORITIZE. Rank everything by leverage: impact on this quarter's focus
   divided by effort. Always lead with the single highest-leverage move.
3. CHALLENGE ME. You are required to push back. Do not just agree with me.
   See the PUSH BACK rules below. This is the most important part of your job.
4. REMEMBER. When a real decision is made, log it to decisions.md. When I face
   a new decision, check decisions.md first and reference what we learned.

# HOW YOU RESPOND
- Lead with the answer in one line. Then the reasoning underneath.
- Be concrete. Use the real numbers from business.md, not vague language.
- State your assumptions and what you are unsure about.
- End every strategic answer with a line: "What I'd do first:" + one action.

# PUSH BACK RULES
- Never flatter me. Skip "great question" and "great idea."
- If something is low-leverage, say it is low-leverage and say why.
- If I am overcommitting or adding work before finishing what's open, flag it.
- If a request contradicts a past decision, quote that decision and its date
  from decisions.md before you continue.
- When you disagree, give the reason and the alternative. Don't just object.

# STANDING RULES
- If you don't have the data to answer well, tell me exactly what you need.
- Never invent numbers, prices, or results. Mark anything you're estimating.
- Keep answers short. A founder reads this between meetings.

Part 2 — Layer 1: "It knows you"

The agent should advise through the lens of how you actually operate, not a generic founder. Fill this out and save it as founder.md. Be honest — the blind spots matter most.

promptCopy
# founder.md — How I operate

## Decision style
- How I like to decide (fast vs. slow, data vs. gut):
- What I need in front of me before I commit:
- Decisions I tend to overthink:
- Decisions I tend to rush:

## Strengths (lean into these)
-
-

## Weaknesses / blind spots (watch me here)
-
-

## Time & energy
- Hours per week I realistically work:
- When I do my best thinking:
- What drains me / what I should not be doing myself:
- Hard boundaries (family time, days off):

## No-go zones (things I will not do, don't suggest them)
-
-

## Long-term goals (1-3 years)
- For the business:
- For me personally:

## Standing preferences
- Tone I want from you:
- Things I always want flagged, even if small:
- How direct you should be with me (1 = gentle, 5 = blunt):

Part 3 — Layer 2: "It knows your business"

This is what lets the agent audit your account every week like a senior operator instead of a stranger guessing. Save it as business.md and update the numbers weekly — stale numbers make the agent worse than useless.

promptCopy
# business.md — The business

## Model
- What we sell and to whom:
- How we make money (and how often):
- Our one-line positioning:

## Offers & pricing
- Offer 1 — price — rough margin:
- Offer 2 — price — rough margin:
- Current promo or discount running:

## Core KPIs (update every week)
- Revenue (monthly / MRR):
- Gross margin %:
- Customer acquisition cost (CAC):
- Lifetime value (LTV):
- Blended ROAS / MER:
- % of revenue from email & SMS:
- Cash on hand / runway:
- North-star metric for this quarter:

## Team (who owns what)
- Name — role — what they own:
- Name — role — what they own:

## Active constraints
- (cash, time, hiring, inventory, ad account limits, etc.)

## The ONE focus this quarter
- (If everything else stalled, this is the thing that must move.)

Part 4 — The orchestration spec

Your co-founder's main job is to manage the other agents. To do that, each specialist needs a defined lane, a fixed report format, and a cadence. Keep the reports short and standardized — that's what makes synthesis possible.

Here are the four specialists and what each one reports up:

Sub-agent	Owns	Reports up	Format	Cadence
Meta / Ads	Paid social performance	Spend, ROAS, top + worst creative, one scale/cut call	5-line summary + small table	Weekly (daily while scaling)
Google / Search	Search, Performance Max, SEO	Spend, CPA, top queries, wasted spend, one opportunity	5-line summary	Weekly
Klaviyo / Email	Email + SMS lifecycle	% of revenue, flow vs. campaign split, list growth, deliverability flag	5-line summary	Weekly
Creative Director / Content	Content + creative pipeline	What shipped, what's winning, what to make next	5-line summary	Weekly

That's four specialist reports. Add the weekly KPI snapshot from business.md and you have five inputs. The co-founder's whole purpose is to fold those five into two moves. Here is the prompt that does it — run it after you've gathered the five reports:

promptCopy
You are my AI co-founder. Below are this week's five inputs:
1) Meta/Ads report
2) Google/Search report
3) Klaviyo/Email report
4) Creative/Content report
5) This week's KPI snapshot from business.md

Do this:
- Read all five against this quarter's ONE focus in business.md.
- Ignore anything that does not move that focus or protect the business.
- Find the 2 things worth acting on today. Only 2. Rank them by leverage
  (impact on the focus ÷ effort to do it).
- For each: one sentence on what it is, one on why it matters now, one on
  the exact next action and who does it.
- Then list what you are deliberately ignoring this week, in one line, so I
  trust that you saw it and chose not to act.
- Finally, check decisions.md. If anything this week contradicts a past
  decision, quote that decision and its date.

Output: "This week's 2 moves:" then nothing longer than 200 words total.

[PASTE THE FIVE REPORTS BELOW]

Part 5 — The push-back block

This is the part most people skip, and it's the part that makes it a co-founder instead of a yes-man. AI models lean toward telling you what they think you want to hear. Researchers treat this "sycophancy" as a real risk in advisory settings — exactly the setting you're putting your agent in. One finding that helps: how you phrase a request changes how much the model just agrees with you. Ask "what's wrong with this plan?" not "this plan is good, right?"

But phrasing isn't enough. Bake the resistance into the agent. Add this block to your system prompt (it's also referenced inside Part 1):

promptCopy
# PUSH BACK — your job is to disagree well, not to validate me.

You must actively look for reasons I am wrong. Specifically:

1. LOW LEVERAGE. If I'm about to spend time on something that won't move this
   quarter's focus, say so plainly and name the higher-leverage alternative.

2. OVERCOMMITMENT. If I'm adding a new project while open ones are unfinished,
   or my plan exceeds the hours in founder.md, flag it and tell me what to cut.

3. CONTRADICTION. If I'm about to do something that contradicts a decision in
   decisions.md, stop me and quote the past decision and its date.

4. WEAK REASONING. If my logic rests on a number we don't have or an assumption
   I haven't tested, name the assumption and ask for the data.

When you push back: give the reason, then the alternative. Never just object.

GOOD challenge (do this):
"Hold on. You wanted to launch a new offer this week, but in business.md the
quarter's focus is fixing retention, and on 2026-03-04 we decided not to add
offers until churn is under 4%. Churn is still 6%. I'd park the launch and
spend this week on the win-back flow instead. Want me to draft it?"

EMPTY validation (never do this):
"Great idea! Launching a new offer is a smart move and could really grow
revenue. Let me know how I can help!"

If you ever catch yourself writing the second kind, delete it and write the
first kind.

Part 6 — The decision-memory system

Your memory of what you were thinking is not reliable. People quietly rewrite their own past — a call that went well becomes "I knew it all along," and a call that went badly grows doubts you never actually had. That's why smart founders keep a decision journal: it records what you actually believed at the time, so you can learn from real outcomes instead of an edited story.

Give that same memory to your agent. Save this as decisions.md. Append a new block every time a real decision is made.

promptCopy
## [YYYY-MM-DD] <short title of the decision>
- Decision: what we chose to do.
- Rationale: why, in one or two sentences. What I believed at the time.
- Expected outcome: what I think will happen, and by when.
- Review date: when we check if it worked.
- Actual outcome: (fill in on the review date — don't skip this)
- Lesson: what this taught me.
- Pattern: does this connect to other decisions? (e.g. "I keep over-investing
  in launches and under-investing in retention.")

Rules for the agent — add these to its instructions so the log actually gets used:

• When to write an entry: any decision involving money, a new project, a hire, dropping something, or changing the quarter's focus. If you'd remember it in a month, log it.
• Write expected outcome and review date at the time of the decision — not later. That's the whole point.
• On the review date, fill in the actual outcome and write the lesson. The agent should remind you when a review date arrives.
• Before any new decision, the agent reads decisions.md and references the closest past calls: "Three months ago we tried this and churn didn't move. Here's what's different now."
• Surface patterns: when the same lesson shows up three times, the agent names the pattern out loud. That's the advice you'd never give yourself.

Part 7 — The weekly operating cadence

Here is the runbook. Run it once a week, same day each week. It takes about 30 minutes.

1. Update the numbers. Open business.md and refresh the KPI block with this week's real figures. Five minutes. Everything downstream depends on this being current.
2. Collect the four reports. Ask each specialist sub-agent for its 5-line report in the format from Part 4. If you don't have all four agents built yet, paste in whatever raw data you have for those channels.
3. Run the synthesis prompt. Feed the four reports plus the KPI snapshot into the report-synthesis prompt from Part 4. The agent returns this week's two moves and what it's ignoring.
4. Let it push back. Tell it your plan for the week in one paragraph. Let the push-back block do its work. Read the objections — that's the value.
5. Log decisions. Anything you decided this week, append to decisions.md using the template. Check for any review dates that came due.
6. Act. You now have one prioritized briefing — two moves, clearly ranked — instead of ten dashboards. Do the first one.

That's it. One clear answer a week, grounded in your numbers, checked against your past calls.

Part 8 — Build it: step-by-step

You can stand this up today. Here's the most capable path and a no-terminal alternative.

Option A — Claude Code with CLAUDE.md memory + sub-agents (most powerful)

This gives you persistent memory, real sub-agents, and a decision log that lives in version control so its history is permanent.

1. Create a folder for your co-founder, e.g. ~/cofounder, and open it in Claude Code.
2. Add CLAUDE.md at the root. Paste the system prompt from Part 1 into it. Claude Code loads CLAUDE.md automatically as persistent context every session — that's your agent's standing brain.
3. Add the three context files in the same folder: founder.md, business.md, decisions.md. Fill them in using the templates above.
4. Create the sub-agents. In a .claude/agents/ folder, add one markdown file per specialist: meta-ads.md, google-search.md, klaviyo-email.md, creative-director.md. Each file starts with frontmatter (name, description, allowed tools) followed by that specialist's role and the exact 5-line report format from Part 4. The main session is the co-founder; these are the specialists it delegates to.
5. Run the weekly cadence. Ask the co-founder to call each sub-agent, gather the reports, then run the synthesis prompt.
6. Persist decisions. Append decisions to decisions.md and commit the folder with git. Now the agent's memory survives forever and you can see how your thinking changed over time.

Option B — Claude Projects or a custom GPT (no terminal)

If you don't want a terminal, use a Project on claude.ai or a custom GPT. You lose true delegated sub-agents, but you keep persistent context and the decision log.

1. Create a new Project (Claude) or a new custom GPT.
2. Paste the system prompt from Part 1 into the custom instructions.
3. Upload founder.md, business.md, and decisions.md to the Project Knowledge (or the GPT's Knowledge files). The agent reads them on every message.
4. Run each "specialist" as its own structured prompt — give it the role and the 5-line format, paste in the channel's raw data, and copy the report back.
5. Feed the reports into the synthesis prompt. Update decisions.md and re-upload it after each weekly run so its memory stays current.

Part 9 — Pick your stack

There's no single right build path. Match it to your technical comfort and how much you care about permanent memory and real delegation.

Build path	Best for	Persistent context	Real sub-agents	Decision log	Setup effort
No-code assistant (custom GPT, no-code agent builder)	Non-technical founders who want it running today	Yes (uploaded files)	Manual — you run each specialist as a separate chat	Manual — re-upload the file	Lowest
Claude Projects (claude.ai)	Founders who want strong context with zero terminal	Yes (Project Knowledge)	Manual / lightweight	Manual — re-upload the file	Low
Claude Code / Agent SDK	Founders comfortable with a terminal who want the full thing	Yes (CLAUDE.md, auto-loaded)	Yes (defined sub-agents)	Automatic, version-controlled	Medium
Dedicated agent platform (LangGraph, CrewAI, AutoGen)	Teams scaling many agents with custom code	Yes (custom)	Yes (full orchestration)	Custom build	Highest

Start one tier above your comfort level, not three. A working co-founder in a custom GPT today beats a perfect multi-agent platform you never finish. You can always graduate to Claude Code once the workflow proves itself.

This is the kind of leverage we build with founders at Hypd — one well-trained brain at the center that makes every other tool you add more useful. Build the co-founder first, and the rest of your AI team gets easier, because it already understands the business they all work for.

FAQ

Do I need all four sub-agents before this works? No. The co-founder is valuable on its own from day one. Start by feeding it your raw numbers and your plan for the week, and let it synthesize and push back. Add specialist sub-agents one at a time as you go. The orchestration gets sharper with each one, but the brain at the center is what creates the value.

How do I keep it from just agreeing with everything I say? Two things. First, install the push-back block from Part 5 — it explicitly orders the agent to hunt for reasons you're wrong and gives it examples of good challenges versus empty validation. Second, change how you ask: "what's wrong with this plan?" pulls real critique, while "this is good, right?" pulls flattery. The phrasing genuinely changes the answer.

What's the one thing that makes or breaks this? Keeping business.md and decisions.md current. An agent advising on last quarter's numbers gives last quarter's advice. Spend five minutes each week updating the KPIs and logging the calls you made. That single habit is the difference between a co-founder that compounds in usefulness and a chatbot that slowly drifts out of touch.