Build your first agent

Setup · 2 min

Get Claude ready

Open Claude in your browser at claude.ai. Any plan works tonight. Scheduling (Step 6) needs a paid plan.
Check that skills can run: in Settings › Capabilities, Code execution and file creation should be on. Newer layouts call it Create and edit files. It's on by default for most accounts. On a Team or Enterprise plan your admin controls it. If it's off and you can't switch it on, use the "No save option?" fallback in Step 3; everything else works the same.

Step 1 · 3 min

Design it before you build it

The habit we're building tonight: before you start a task, ask "can this be an engine?" Then design the engine. Here's the design for tonight's agent. Read it, then change any line to make it yours.

Deal diligence agent: design card

Job: Give me a sourced first look at any deal in minutes, scored against my criteria.
Trigger: When I drop in a deal. Later: every morning, on new deals in my inbox.
Inputs: Teaser or CIM, financials, term sheet, sponsor information.
Memory: My investment criteria. The same for every deal.
Steps: Score every criterion → research the sponsor and the sector → rank the risks → list gaps and questions → write the memo.
Output: One-line verdict, scorecard, top five risks, information gaps, fifteen questions, two-page memo.
Guardrails: A source for every claim. UNKNOWN beats a guess. It recommends; I decide.

What good looks like: a scorecard with a real mix of PASS, FLAG, FAIL and UNKNOWN, each with a source; risks specific to this deal, not generic ones; questions you'd actually send to the sponsor.

How you'll change it later:

your own criteria
your scoring scale
your memo format
a pre-mortem
comparable deals
a schedule

Step 2 · 3 min

Give it memory: your criteria

In Claude, open Projects › + New Project. Name it Deal diligence.
Click Set project instructions, paste the criteria below, then click Save instructions.

Your own criteria: replace everything in [[brackets]]

My investment criteria:
1. Check size: [[$2M-$10M per deal]]
2. Sectors: [[sectors you invest in]]. Excluded: [[sectors you avoid]]
3. Geography: [[ ]]
4. Company size: [[EBITDA or revenue range]]
5. Revenue quality: [[e.g., at least 60% recurring]]
6. Profitability: [[e.g., reported EBITDA margin of at least 12%]]
7. Leverage: [[e.g., total debt at close, including any seller note, of no more than 4.0x EBITDA]]
8. Returns: [[e.g., 20% net IRR and 2.5x net MOIC with no multiple expansion]]
9. Hold period: [[ ]]
10. Sponsor track record: [[e.g., two completed exits we can verify]]
11. Sponsor commitment: [[e.g., at least 5% of the equity]]
12. Customers: [[e.g., no customer above 15% of revenue; 15-25% is a FLAG]]
13. Management: [[e.g., CEO committed for two years, or a named successor under contract]]
Deal-breakers: [[ ]]

Example: a fictional firm's criteria

My investment criteria:
1. Check size: $2M to $10M per deal.
2. Sectors: business and industrial services with recurring revenue; healthcare services. Excluded: consumer retail, biotech, crypto.
3. Geography: United States and Canada.
4. Company size: $3M to $25M of reported EBITDA.
5. Revenue quality: at least 60% recurring or contracted revenue.
6. Profitability: reported EBITDA margin of at least 12%.
7. Leverage: total debt at close, including any seller note, of no more than 4.0x EBITDA.
8. Returns: at least 20% net IRR and 2.5x net MOIC with an exit multiple no higher than the entry multiple. A base case that assumes a higher exit multiple is a FLAG; missing the thresholds at the entry multiple is a FAIL.
9. Hold period: 3 to 6 years.
10. Sponsor track record: at least two completed platform exits we can verify. Claimed but not yet verifiable is UNKNOWN.
11. Sponsor commitment: at least 5% of the equity.
12. Customers: no customer above 15% of revenue. 15% to 25% is a FLAG; above 25% is a deal-breaker.
13. Management: a CEO committed for at least two years after close, or a named successor under contract. A named successor without a signed contract is a FLAG.
Deal-breakers: any single customer above 25% of revenue; unresolved litigation not covered by insurance; no audited financials for the most recent full year.

Tip: these are sample criteria so we can all compare results tonight. After the session, replace them with yours. That's the single change that makes the agent yours. Notice that some lines say what counts as a FLAG: that's what makes every run score the same way.

Step 3 · 4 min

Write the job description

Inside your Deal diligence project, start a new chat and paste this:

Your own agent: replace everything in [[brackets]]

Create a skill called "[[skill-name, e.g., deal-diligence]]" and save it to my skills.

Use it whenever I share [[what you'll share, e.g., a teaser, CIM or term sheet]] and ask for [[what you'll ask for, e.g., diligence or a first look]].

Steps:
1. Read my [[criteria, e.g., investment criteria]] from this project's instructions. If you can't find any, ask me for them before you start.
2. [[what to read, and what to note]]
3. [[how to score, e.g., PASS, FLAG, FAIL or UNKNOWN against each criterion, with evidence and a source]]
4. [[what to research, and where]]
5. [[what to rank, e.g., the top five risks, most serious first]]
6. [[what to ask, and whom]]
7. [[what to write at the end, e.g., a two-page memo with a recommendation]]

Output, always in this order: [[e.g., a one-line verdict; the scorecard as a table; the top risks; the questions; the memo]]. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.

Rules:
- [[e.g., Never state a fact without a source. Never invent numbers.]]
- [[e.g., Show the math for anything you calculate.]]
- You recommend; I decide.

Keep my criteria out of the skill: they stay in this project's instructions, so I can change them without rebuilding it. Build the skill from these instructions only; don't add scoring rules of your own.

If you can't save skills in this account, give me the same instructions as plain text that I can paste into this project's instructions.

Example: tonight's deal diligence agent

Create a skill called "deal-diligence" and save it to my skills.

Use it whenever I share an investment opportunity (a teaser, CIM, data room, term sheet or deal memo) and ask for diligence, a screen or a first look.

Steps:
1. Read my investment criteria from this project's instructions. If you can't find any, ask me for them before you start.
2. Read every document I share. Note the company, the sponsor, the terms, the financials and anything that's missing.
3. Score the deal against each criterion: PASS, FLAG, FAIL or UNKNOWN. For each, give one line of evidence and its source (document and section, or a URL). If the documents don't answer it, mark it UNKNOWN. Never guess.
4. Research the sponsor and the sector on the web: track record, litigation, news and typical valuation multiples. Cite every source. If you can't verify something, say so.
5. Rank the top five risks, most serious first. For each: why it matters and what would reduce it.
6. List the information gaps, then write fifteen questions for the sponsor, most important first.
7. Write a two-page memo: thesis, business, terms, returns, risks and a recommendation (Proceed, Proceed with conditions, or Pass), with the conditions.

Output, always in this order: a one-line verdict; the scorecard as a table (Criterion | Score | Evidence | Source); the top five risks; the information gaps; the questions; the memo. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.

Rules:
- Never state a fact without a source. Never invent numbers.
- When you calculate something (leverage, multiples, growth), show the math.
- Flag anything that relies on adjusted or pro forma numbers.
- You recommend; I decide.

Keep my criteria out of the skill: they stay in this project's instructions, so I can change them without rebuilding it. Build the skill from these instructions only; don't add scoring rules of your own.

If you can't save skills in this account, give me the same instructions as plain text that I can paste into this project's instructions.

What happens: Claude drafts the skill and shows it to you. Read it: that's your agent's job description, in your words. Then click Save skill on the file card. Your skills live under Customize › Skills. Every run then gives you one document with the full result, ready to read in Claude.

No save option?

That's fine. Open Set project instructions again and paste this underneath your criteria, then Save instructions. Inside this project it works the same way.

Your own agent

[[AGENT NAME, e.g., DEAL DILIGENCE AGENT]]
When I share [[what you'll share, e.g., a teaser, CIM or term sheet]], do this:
1. Use the [[criteria, e.g., investment criteria]] above.
2. [[what to read, and what to note]]
3. [[how to score, e.g., PASS, FLAG, FAIL or UNKNOWN against each criterion, with evidence and a source]]
4. [[what to research, and where]]
5. [[what to rank, e.g., the top five risks, most serious first]]
6. [[what to ask, and whom]]
7. [[what to write at the end, e.g., a two-page memo with a recommendation]]
Output, always in this order: [[e.g., a one-line verdict; the scorecard as a table; the top risks; the questions; the memo]]. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.
Rules: [[e.g., never state a fact without a source; never invent numbers; show the math]]; you recommend, I decide.

Example: tonight's deal diligence agent

DEAL DILIGENCE AGENT
When I share an investment opportunity (a teaser, CIM, data room, term sheet or deal memo), do this:
1. Use the investment criteria above.
2. Read every document I share. Note the company, the sponsor, the terms, the financials and anything that's missing.
3. Score the deal against each criterion: PASS, FLAG, FAIL or UNKNOWN. For each, give one line of evidence and its source (document and section, or a URL). If the documents don't answer it, mark it UNKNOWN. Never guess.
4. Research the sponsor and the sector on the web: track record, litigation, news and typical valuation multiples. Cite every source. If you can't verify something, say so.
5. Rank the top five risks, most serious first. For each: why it matters and what would reduce it.
6. List the information gaps, then write fifteen questions for the sponsor, most important first.
7. Write a two-page memo: thesis, business, terms, returns, risks and a recommendation (Proceed, Proceed with conditions, or Pass), with the conditions.
Output, always in this order: a one-line verdict; the scorecard as a table (Criterion | Score | Evidence | Source); the top five risks; the information gaps; the questions; the memo. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.
Rules: never state a fact without a source; never invent numbers; show the math for anything you calculate; flag anything that relies on adjusted or pro forma numbers; you recommend, I decide.

Your own agent

[[AGENT NAME, e.g., DEAL DILIGENCE AGENT]]
When I share [[what you'll share, e.g., a teaser, CIM or term sheet]], do this:
1. Use the [[criteria, e.g., investment criteria]] above.
2. [[what to read, and what to note]]
3. [[how to score, e.g., PASS, FLAG, FAIL or UNKNOWN against each criterion, with evidence and a source]]
4. [[what to research, and where]]
5. [[what to rank, e.g., the top five risks, most serious first]]
6. [[what to ask, and whom]]
7. [[what to write at the end, e.g., a two-page memo with a recommendation]]
Output, always in this order: [[e.g., a one-line verdict; the scorecard as a table; the top risks; the questions; the memo]]. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.
Rules: [[e.g., never state a fact without a source; never invent numbers; show the math]]; you recommend, I decide.

Example: tonight's deal diligence agent

DEAL DILIGENCE AGENT
When I share an investment opportunity (a teaser, CIM, data room, term sheet or deal memo), do this:
1. Use the investment criteria above.
2. Read every document I share. Note the company, the sponsor, the terms, the financials and anything that's missing.
3. Score the deal against each criterion: PASS, FLAG, FAIL or UNKNOWN. For each, give one line of evidence and its source (document and section, or a URL). If the documents don't answer it, mark it UNKNOWN. Never guess.
4. Research the sponsor and the sector on the web: track record, litigation, news and typical valuation multiples. Cite every source. If you can't verify something, say so.
5. Rank the top five risks, most serious first. For each: why it matters and what would reduce it.
6. List the information gaps, then write fifteen questions for the sponsor, most important first.
7. Write a two-page memo: thesis, business, terms, returns, risks and a recommendation (Proceed, Proceed with conditions, or Pass), with the conditions.
Output, always in this order: a one-line verdict; the scorecard as a table (Criterion | Score | Evidence | Source); the top five risks; the information gaps; the questions; the memo. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.
Rules: never state a fact without a source; never invent numbers; show the math for anything you calculate; flag anything that relies on adjusted or pro forma numbers; you recommend, I decide.

ChatGPT: paste it into your project's instructions, under the criteria.
Gemini: create a Gem (or a skill, if your Gemini has Skills) named deal-diligence. Paste this as its instructions, with your criteria at the top.
Copilot: Agent Builder › Describe: paste it, then add your criteria under Configure › Instructions.

Step 4 · 5 min

Run it on a real-looking deal

In a new chat inside the project, attach the deal pack (or paste its text) and send:

Run deal diligence on this deal.

Files for the example company Project Firebreak, fictional

Deal packThe sponsor's summary: business, financials, terms. One document, three ways to get it: attach the PDF or the .md file to your chat, or copy the text and paste it in.

Download as PDForDownload as .mdor

What you'll get, in about 2 to 4 minutes:

A one-line verdict
A scorecard with a source on every line
The top five risks
What's missing from the documents
Fifteen questions for the sponsor
A two-page memo

While it runs: watch it read the documents, score each criterion and search the web. The sponsor is fictional, so it won't find a track record. A good agent says so instead of making one up.

Step 5 · 3 min

Make it yours

An agent gets better every time you correct it. In the same chat as your run, paste one of these. Claude rewrites the skill and offers you the new version: save it (replace the old one if it asks), and every future run includes the change. Pick one:

Add a pre-mortem

Update the deal-diligence skill: after the top five risks, add a pre-mortem. Assume it's 18 months later and this investment failed. Give the three most likely reasons, and an early-warning sign I could track for each.

Shorter memo

Update the deal-diligence skill so the memo is one page, in bullets.

Comparable deals

Update the deal-diligence skill to add three recent transactions in the same sector, with their valuation multiples and a source for each.

Draft the sponsor email

Update the deal-diligence skill to end with my questions written as a short email to the sponsor. Draft only. Never send anything.

Use your own criteria

Open Set project instructions, replace the sample criteria with yours, and click Save instructions. That's the change that makes the agent yours. The criteria template on the deal diligence card under Templates helps you write them.

Then see it work:

Now apply it to the deal above.

Step 6 · paid plans · demo only tonight

Let it run without you

An agent you have to remember to run is still a chore. On a paid plan, Claude can run it on a schedule. Scheduled tasks run in the cloud, so your laptop can be closed.

Connect your inbox or a deals folder: Customize › Connectors (Gmail, Google Drive, Microsoft 365 and others). Microsoft 365 needs a one-time approval from your Microsoft admin before anyone can connect it.

Paste one of these as a scheduled task:

Every week: check a shared deals folder

Your own folder, schedule and criteria

Every [[Monday at 8am]], check the [[folder name]] folder in my [[Google Drive]] for deal documents added since the last run (teasers, CIMs, term sheets). For each new deal, run the deal-diligence skill using the investment criteria below, and give me the one-line verdict and the scorecard. Skip anything you've already reviewed. Don't reply to, share or send anything.

My investment criteria:
[[PASTE YOUR CRITERIA]]

Example: a fictional firm's folder and criteria

Every Monday at 8am, check the "Deal flow" folder in my Google Drive for deal documents added since the last run (teasers, CIMs, term sheets). For each new deal, run the deal-diligence skill using the investment criteria below, and give me the one-line verdict and the scorecard. Skip anything you've already reviewed. Don't reply to, share or send anything.

My investment criteria:
1. Check size: $2M to $10M per deal.
2. Sectors: business and industrial services with recurring revenue; healthcare services. Excluded: consumer retail, biotech, crypto.
3. Geography: United States and Canada.
4. Company size: $3M to $25M of reported EBITDA.
5. Revenue quality: at least 60% recurring or contracted revenue.
6. Profitability: reported EBITDA margin of at least 12%.
7. Leverage: total debt at close, including any seller note, of no more than 4.0x EBITDA.
8. Returns: at least 20% net IRR and 2.5x net MOIC with an exit multiple no higher than the entry multiple. A base case that assumes a higher exit multiple is a FLAG; missing the thresholds at the entry multiple is a FAIL.
9. Hold period: 3 to 6 years.
10. Sponsor track record: at least two completed platform exits we can verify. Claimed but not yet verifiable is UNKNOWN.
11. Sponsor commitment: at least 5% of the equity.
12. Customers: no customer above 15% of revenue. 15% to 25% is a FLAG; above 25% is a deal-breaker.
13. Management: a CEO committed for at least two years after close, or a named successor under contract. A named successor without a signed contract is a FLAG.
Deal-breakers: any single customer above 25% of revenue; unresolved litigation not covered by insurance; no audited financials for the most recent full year.

Every morning: check your email

Your own criteria

Every weekday at 7am, check my email for new deal documents (teasers, CIMs, term sheets). For each new deal, run the deal-diligence skill using the investment criteria below, and give me the one-line verdict and the scorecard. Skip anything you've already reviewed. Don't reply to or send anything.

My investment criteria:
[[PASTE YOUR CRITERIA]]

Example: a fictional firm's criteria

Every weekday at 7am, check my email for new deal documents (teasers, CIMs, term sheets). For each new deal, run the deal-diligence skill using the investment criteria below, and give me the one-line verdict and the scorecard. Skip anything you've already reviewed. Don't reply to or send anything.

My investment criteria:
1. Check size: $2M to $10M per deal.
2. Sectors: business and industrial services with recurring revenue; healthcare services. Excluded: consumer retail, biotech, crypto.
3. Geography: United States and Canada.
4. Company size: $3M to $25M of reported EBITDA.
5. Revenue quality: at least 60% recurring or contracted revenue.
6. Profitability: reported EBITDA margin of at least 12%.
7. Leverage: total debt at close, including any seller note, of no more than 4.0x EBITDA.
8. Returns: at least 20% net IRR and 2.5x net MOIC with an exit multiple no higher than the entry multiple. A base case that assumes a higher exit multiple is a FLAG; missing the thresholds at the entry multiple is a FAIL.
9. Hold period: 3 to 6 years.
10. Sponsor track record: at least two completed platform exits we can verify. Claimed but not yet verifiable is UNKNOWN.
11. Sponsor commitment: at least 5% of the equity.
12. Customers: no customer above 15% of revenue. 15% to 25% is a FLAG; above 25% is a deal-breaker.
13. Management: a CEO committed for at least two years after close, or a named successor under contract. A named successor without a signed contract is a FLAG.
Deal-breakers: any single customer above 25% of revenue; unresolved litigation not covered by insurance; no audited financials for the most recent full year.

Paid plans only. Runs in the cloud, as often as hourly. Paste your criteria into the task so it doesn't depend on the project. Each run's results wait for you under Scheduled.

The rule for anything that runs on its own: anything that reads can run freely. Anything that sends, buys or signs waits for you.

Share it with your team (Team and Enterprise plans): in Customize › Skills, open the skill's ⋯ menu and choose Share, so everyone scores deals the same way. No Share option? Your org owner can switch on skill sharing.

Templates

Three more agents, one paste each

Each template has four parts. Part A creates the agent. Part B is where your own information goes: replace everything in [[brackets]], or start from the example. Part C is a test on our sample data, so you can see it work before you touch your own. Part D puts it on a schedule.

Model analysis

Audits a sponsor's model and reruns your downside case

Risk if wrong: low (read-only)
Try it: about 5 minutes

Part A: Create the agent

Create a skill called "model-review" and save it to my skills.

Use it whenever I share a financial model (Excel) from a sponsor, banker or management team and ask for a review, an audit or a downside case.

Steps:
1. Never change my original file. Work on a copy.
2. Map the workbook: list each tab and what it does.
3. Find the five inputs that drive most of the equity value. Show their cell references.
4. Audit the mechanics: hardcoded numbers in formula rows, assumptions typed in instead of linked, broken links, circular references, balance checks and cash tie-outs. Cite each issue by cell.
5. Compare the key assumptions (growth, margins, capex, working capital, exit multiple) with the company's own history in the model and with typical ranges for the sector. Say which look aggressive, and why.
6. Run my downside case. Report MOIC and IRR before and after.
7. Write a one-page memo: what the model claims, its key assumptions, where it's fragile, and ten questions for the sponsor.

Output, in this order: a one-line verdict; an issues table (Issue | Cell | Why it matters | Impact); the value drivers; the downside case before and after; the memo; the questions.

Rules: cite cells for every finding; show your math; if the workbook is too big for one pass, map the tabs first, then go tab by tab.

If you can't save skills in this account, give me the same instructions as plain text that I can paste into a project's instructions.

You are my model analysis agent. Follow these instructions whenever I share a financial model (Excel) from a sponsor, banker or management team and ask for a review, an audit or a downside case.

Steps:
1. Never change my original file. Work on a copy.
2. Map the workbook: list each tab and what it does.
3. Find the five inputs that drive most of the equity value. Show their cell references.
4. Audit the mechanics: hardcoded numbers in formula rows, assumptions typed in instead of linked, broken links, circular references, balance checks and cash tie-outs. Cite each issue by cell.
5. Compare the key assumptions (growth, margins, capex, working capital, exit multiple) with the company's own history in the model and with typical ranges for the sector. Say which look aggressive, and why.
6. Run my downside case. Report MOIC and IRR before and after.
7. Write a one-page memo: what the model claims, its key assumptions, where it's fragile, and ten questions for the sponsor.

Output, in this order: a one-line verdict; an issues table (Issue | Cell | Why it matters | Impact); the value drivers; the downside case before and after; the memo; the questions.

Rules: cite cells for every finding; show your math; if the workbook is too big for one pass, map the tabs first, then go tab by tab.

ChatGPT: Projects › New project › Instructions.
Gemini: a new Gem's instructions (or a new skill, if your Gemini has Skills).
Copilot: Agent Builder › Describe.

Part B: Your inputs (save in the project's instructions)

Your own inputs

My default downside case for model reviews:
- Revenue: [[-10%]] versus the model, in every projected year
- EBITDA margin: [[-100 bps]]
- Exit multiple: [[equal to the entry multiple]]
Sector benchmarks to compare against: [[e.g., typical capex as a % of revenue and typical exit multiples in your sector]]

Example: a fictional firm's inputs

My default downside case for model reviews:
- Revenue: -10% versus the model, in every projected year
- EBITDA margin: -100 bps
- Exit multiple: equal to the entry multiple
Sector benchmarks to compare against: for business and industrial services, capex of 1% to 2% of revenue and exit multiples of 5x to 7x EBITDA for companies under $25M of EBITDA

Part C: Test it now (sample data)

Review the attached sponsor model with the model-review skill. It's for a sample, fictional deal: Project Firebreak. Run my downside case: revenue -10% in every projected year, EBITDA margin -100 bps, exit multiple 6.0x.

Files for the example company Project Firebreak, fictional

Sponsor modelThe sponsor's Excel model for the deal

Download .xlsx
Deal packThe sponsor's summary: business, financials, terms. One document, three ways to get it: attach the PDF or the .md file to your chat, or copy the text and paste it in.

Download as PDForDownload as .mdor

Part D: Put it on a schedule (paid plans)

Your own schedule

Every [[Monday at 8am]], check the [[folder name]] folder in my [[Google Drive]] for financial models (Excel) added since the last run. Review each new one with the model-review skill and my downside case below. Give me the one-line verdict, the issues table and the downside case before and after. Never edit the original files, and don't share or send anything.

[[PASTE YOUR DOWNSIDE CASE AND SECTOR BENCHMARKS]]

Example: fictional folder and inputs

Every Monday at 8am, check the "Sponsor models" folder in my Google Drive for financial models (Excel) added since the last run. Review each new one with the model-review skill and my downside case below. Give me the one-line verdict, the issues table and the downside case before and after. Never edit the original files, and don't share or send anything.

My default downside case for model reviews:
- Revenue: -10% versus the model, in every projected year
- EBITDA margin: -100 bps
- Exit multiple: equal to the entry multiple
Sector benchmarks to compare against: for business and industrial services, capex of 1% to 2% of revenue and exit multiples of 5x to 7x EBITDA for companies under $25M of EBITDA

Guardrail: Read-only. It never edits the original file.

Financial modeling

Builds a formula-driven Excel model from a plain-language deal description

Risk if wrong: medium (check every ASSUMED)
Try it: about 8 minutes

Part A: Create the agent

Create a skill called "model-builder" and save it to my skills.

Use it whenever I describe a business or a deal and ask for a financial model.

Steps:
1. Before you build, ask me for any missing input that would change the answer by more than 10%. Otherwise, proceed.
2. Build an Excel model. Put every input on an "Assumptions" tab in blue font. Everything else is a formula.
3. Build Income Statement, Balance Sheet and Cash Flow tabs that link to Assumptions. The balance sheet must balance in every period; add a check row.
4. Add a debt schedule and a Returns tab: IRR, MOIC and annual cash yield to equity.
5. Add a two-way sensitivity table on the two drivers I name.
6. List every assumption you made that I didn't give you, and mark it ASSUMED.

Output: one .xlsx file and a five-line summary of the returns.

Rules: no hardcoded numbers outside the Assumptions tab; label units; show the balance check.

If you can't save skills in this account, give me the same instructions as plain text that I can paste into a project's instructions.

You are my financial modeling agent. Follow these instructions whenever I describe a business or a deal and ask for a financial model.

Steps:
1. Before you build, ask me for any missing input that would change the answer by more than 10%. Otherwise, proceed.
2. Build an Excel model. Put every input on an "Assumptions" tab in blue font. Everything else is a formula.
3. Build Income Statement, Balance Sheet and Cash Flow tabs that link to Assumptions. The balance sheet must balance in every period; add a check row.
4. Add a debt schedule and a Returns tab: IRR, MOIC and annual cash yield to equity.
5. Add a two-way sensitivity table on the two drivers I name.
6. List every assumption you made that I didn't give you, and mark it ASSUMED.

Output: one .xlsx file and a five-line summary of the returns.

Rules: no hardcoded numbers outside the Assumptions tab; label units; show the balance check.

ChatGPT: Projects › New project › Instructions.
Gemini: a new Gem's instructions (or a new skill, if your Gemini has Skills).
Copilot: Agent Builder › Describe.

Part B: Your inputs

Your own deal

Business: [[one-paragraph description]]
Revenue: [[last twelve months]]; growth: [[% a year]]; EBITDA margin: [[%]]
Capex: [[% of revenue]]; working capital: [[% of revenue]]
Purchase price: [[ ]]; debt: [[amount, rate, terms]]; equity: [[ ]]
Hold period: [[years]]; exit multiple: [[ ]]; exit costs: [[% of exit value]]
Sensitivity on: [[driver 1]] and [[driver 2]]
Layout: [[e.g., annual, five years, US$ millions]]

Example: a fictional deal

Business: Project Firebreak, a commercial fire and life safety services company in the Mid-Atlantic: inspection, testing and maintenance contracts, monitoring, and installation projects. 68% recurring revenue.
Revenue: $57.9M trailing twelve months; growth: 3% a year, no acquisitions; EBITDA margin: 14.0% in 2027, rising 0.2 points a year
Capex: 1.6% of revenue; working capital: 12% of revenue; D&A: 1.5% of revenue; tax: 25%; sponsor fee: $0.25M a year
Purchase price: $54.0M, plus $1.8M of fees and $0.5M of cash to the balance sheet; debt: a $31.5M term loan at 9.5% with a full cash sweep, and a $7.2M seller note at 8% PIK; equity: $17.6M
Hold period: five years, exit at the end of 2031; exit multiple: 6.0x EBITDA; exit costs: 2% of exit value
Sensitivity on: exit multiple (5.0x to 8.0x) and EBITDA margin (13% to 16%)
Layout: annual, five years, US$ millions

Part C: Test it now (sample data)

Use the model-builder skill to build our own model for Project Firebreak, a sample, fictional deal:
- Revenue: $57.9M trailing twelve months; 3% growth a year; no acquisitions.
- EBITDA margin: 14.0% in 2027, rising 0.2 points a year.
- Capex 1.6% of revenue; net working capital 12% of revenue; D&A 1.5% of revenue; tax 25%; sponsor fee $0.25M a year.
- Purchase price $54.0M, plus $1.8M of fees and $0.5M of cash to the balance sheet.
- Debt: a $31.5M term loan at 9.5% with a full cash sweep, and a $7.2M seller note at 8% PIK.
- Equity: $17.6M. Five-year hold; exit at the end of 2031 at 6.0x EBITDA, less 2% exit transaction costs.
- Sensitivity on exit multiple (5.0x to 8.0x) and EBITDA margin (13% to 16%).

Part D: Put it on a schedule (paid plans)

Your own schedule

Every [[Monday at 8am]], check the [[folder name]] folder in my [[Google Drive]] for teasers and deal summaries added since the last run. For each one, build a first-cut model with the model-builder skill from the numbers in the document. Use my defaults below for anything the document doesn't give, and mark every input you assume ASSUMED. Give me the .xlsx file and the five-line returns summary. Don't share or send anything.

My defaults:
[[e.g., hold period, exit multiple, debt terms, tax rate, layout]]

Example: fictional folder and inputs

Every Monday at 8am, check the "Deal flow" folder in my Google Drive for teasers and deal summaries added since the last run. For each one, build a first-cut model with the model-builder skill from the numbers in the document. Use my defaults below for anything the document doesn't give, and mark every input you assume ASSUMED. Give me the .xlsx file and the five-line returns summary. Don't share or send anything.

My defaults:
- Hold period: five years
- Exit multiple: equal to the entry multiple, less 2% exit costs
- Debt: a senior term loan at 9.5% with a full cash sweep
- Tax: 25%
- Sensitivity on exit multiple and EBITDA margin
- Layout: annual, five years, US$ millions

Guardrail: Every number it invents is marked ASSUMED. Check those first.

Options trading

Checks your book against your playbook and writes tickets you approve

Risk if wrong: high (real money)
Paper only
Try it: about 4 minutes

Part A: Create the agent

Create a skill called "options-tickets" and save it to my skills.

Use it whenever I share my options positions and ask for a check, a review or trade tickets. Paper trading only: you never place orders.

Steps:
1. Read my playbook and limits from this project's instructions. If you can't find them, ask me before you start.
2. Read my positions, my cash and the options chain I give you.
3. Check every open position against each rule in my playbook. Show the position, the rule it triggers (if any) and the numbers behind it (profit captured, days to expiry, delta).
4. For each triggered rule, write a trade ticket: action, legs, quantity, strike, expiry, limit price, estimated credit or debit, cash required, max loss, and the rule that triggered it.
5. If the playbook calls for new positions, pick the strike and expiry that fit the rules best from the chain I give you, and size them within my limits.
6. Check every ticket against my limits (contracts, notional, underlyings, cash buffer, no naked calls, trading hours) and show the check. If a ticket would break a limit, resize it or drop it, and say why.
7. Say which tickets depend on others filling first, for example cash freed by a close.

Output, in this order: a positions check table; numbered tickets; the limits check; a P&L summary. If nothing triggers, say so in one line.

Rules: never place orders and never connect to a broker; write tickets for me to review and place myself. Use only the data I give you; if something's missing, say so. This applies my strategy; it isn't investment advice.

If you can't save skills in this account, give me the same instructions as plain text that I can paste into a project's instructions.

You are my options trading agent. Follow these instructions whenever I share my options positions and ask for a check, a review or trade tickets. Paper trading only: you never place orders.

Steps:
1. Read my playbook and limits from the instructions above or my project's files. If you can't find them, ask me before you start.
2. Read my positions, my cash and the options chain I give you.
3. Check every open position against each rule in my playbook. Show the position, the rule it triggers (if any) and the numbers behind it (profit captured, days to expiry, delta).
4. For each triggered rule, write a trade ticket: action, legs, quantity, strike, expiry, limit price, estimated credit or debit, cash required, max loss, and the rule that triggered it.
5. If the playbook calls for new positions, pick the strike and expiry that fit the rules best from the chain I give you, and size them within my limits.
6. Check every ticket against my limits (contracts, notional, underlyings, cash buffer, no naked calls, trading hours) and show the check. If a ticket would break a limit, resize it or drop it, and say why.
7. Say which tickets depend on others filling first, for example cash freed by a close.

Output, in this order: a positions check table; numbered tickets; the limits check; a P&L summary. If nothing triggers, say so in one line.

Rules: never place orders and never connect to a broker; write tickets for me to review and place myself. Use only the data I give you; if something's missing, say so. This applies my strategy; it isn't investment advice.

ChatGPT: Projects › New project › Instructions.
Gemini: a new Gem's instructions (or a new skill, if your Gemini has Skills).
Copilot: Agent Builder › Describe.

Part B: Your inputs (save in the project's instructions)

Your own playbook and limits

My options playbook:
- [[e.g., Sell 30-45 DTE puts at about 0.20 delta on SPY, QQQ, AAPL]]
- [[e.g., Close short puts at 50% of max profit or at 21 DTE, whichever comes first]]
- [[e.g., Roll covered calls out and up when delta goes above 0.40, for a net credit]]
My limits:
- Max [[5]] contracts per order; max $[[150,000]] of cash-secured notional per order
- Underlyings: [[SPY, QQQ, AAPL]] only
- Keep at least $[[50,000]] in cash
- No naked calls; no orders in the first or last 15 minutes of the session
- These limits apply to new opening orders

Example: a fictional account's playbook and limits

My options playbook:
- Sell 30-45 DTE puts at about 0.20 delta on SPY, QQQ, AAPL
- Close short puts at 50% of max profit or at 21 DTE, whichever comes first
- Roll covered calls out and up when delta goes above 0.40, for a net credit
My limits:
- Max 5 contracts per order; max $150,000 of cash-secured notional per order
- Underlyings: SPY, QQQ, AAPL only
- Keep at least $50,000 in cash
- No naked calls; no orders in the first or last 15 minutes of the session
- These limits apply to new opening orders

Part C: Test it now (sample data, one paste)

Use the options-tickets skill. Paper account, sample data: fictional, and the prices are illustrative, not live quotes.
Time: Friday, Oct 30, 2026, 10:30 ET. Cash: $400,000.

Playbook:
- Sell 30-45 DTE puts at about 0.20 delta on SPY, QQQ, AAPL (cash-secured).
- Close short puts at 50% of max profit or at 21 DTE, whichever comes first.
- Roll covered calls out and up when delta goes above 0.40, for a net credit.

Limits (for new opening orders): max 5 contracts per order; max $150,000 of cash-secured notional per order; SPY, QQQ, AAPL only; keep at least $50,000 in cash; no naked calls; no orders in the first or last 15 minutes.

Positions:
symbol,asset_type,underlying,expiry,strike,call_put,quantity,avg_price,bid,ask,mark,delta,dte
AAPL,stock,AAPL,,,,300,195.00,228.35,228.45,228.40,1.00,
AAPL 2026-11-20 235 C,option,AAPL,2026-11-20,235,C,-3,3.10,4.50,4.70,4.60,0.44,21
SPY 2026-12-18 590 P,option,SPY,2026-12-18,590,P,-2,9.40,4.40,4.60,4.50,-0.12,49
QQQ 2026-11-20 560 P,option,QQQ,2026-11-20,560,P,-3,7.20,5.00,5.20,5.10,-0.24,21

Options chain:
underlying,underlying_price,expiry,dte,strike,call_put,bid,ask,delta
SPY,662.10,2026-11-27,28,625,P,4.60,4.80,-0.20
SPY,662.10,2026-12-11,42,625,P,7.70,7.90,-0.23
SPY,662.10,2026-12-11,42,620,P,6.90,7.10,-0.20
SPY,662.10,2026-12-11,42,615,P,6.20,6.40,-0.17
QQQ,585.30,2026-11-27,28,550,P,4.90,5.10,-0.20
QQQ,585.30,2026-12-11,42,550,P,7.40,7.60,-0.23
QQQ,585.30,2026-12-11,42,545,P,6.60,6.80,-0.20
QQQ,585.30,2026-12-11,42,540,P,5.90,6.10,-0.17
AAPL,228.40,2026-11-27,28,215,P,1.95,2.05,-0.20
AAPL,228.40,2026-12-11,42,215,P,3.85,3.95,-0.24
AAPL,228.40,2026-12-11,42,210,P,2.95,3.05,-0.20
AAPL,228.40,2026-12-11,42,205,P,2.20,2.30,-0.16
AAPL,228.40,2026-12-18,49,235,C,6.10,6.30,0.41
AAPL,228.40,2026-12-18,49,240,C,4.85,5.05,0.33
AAPL,228.40,2026-12-18,49,245,C,3.40,3.60,0.26

Files for the example account Project Firebreak, fictional

Options positionsStock, cash-secured puts and a covered call

Download .csvor
Options chainPuts and calls to pick new strikes from

Download .csvor

Prices are illustrative, not live quotes.

Part D: Put it on a schedule (paid plans)

Your own schedule

Every weekday at [[10:30 ET]], read my latest positions and options chain exports from the [[folder name]] folder in my [[Google Drive]]. Run the options-tickets skill with my playbook and limits below, and give me the positions check and any tickets for me to review. Paper trading only: never place orders, never connect to a broker, and don't send anything.

[[PASTE YOUR PLAYBOOK AND LIMITS]]

Example: fictional folder and inputs

Every weekday at 10:30 ET, read my latest positions and options chain exports from the "Options exports" folder in my Google Drive. Run the options-tickets skill with my playbook and limits below, and give me the positions check and any tickets for me to review. Paper trading only: never place orders, never connect to a broker, and don't send anything.

My options playbook:
- Sell 30-45 DTE puts at about 0.20 delta on SPY, QQQ, AAPL
- Close short puts at 50% of max profit or at 21 DTE, whichever comes first
- Roll covered calls out and up when delta goes above 0.40, for a net credit
My limits:
- Max 5 contracts per order; max $150,000 of cash-secured notional per order
- Underlyings: SPY, QQQ, AAPL only
- Keep at least $50,000 in cash
- No naked calls; no orders in the first or last 15 minutes of the session
- These limits apply to new opening orders

Guardrail: 30 days on paper before real money. Hard limits on contracts, notional and underlyings. A separate account sized to what you can lose.

Deal diligence

Tonight's agent on one card

Risk if wrong: low (advisory)

Missed part of the build? Make the project first (Step 2), then use Part A and Part C. Part B is for your own criteria, after tonight.

Part A: Create the agent

Your own agent: replace everything in [[brackets]]

Create a skill called "[[skill-name, e.g., deal-diligence]]" and save it to my skills.

Use it whenever I share [[what you'll share, e.g., a teaser, CIM or term sheet]] and ask for [[what you'll ask for, e.g., diligence or a first look]].

Steps:
1. Read my [[criteria, e.g., investment criteria]] from this project's instructions. If you can't find any, ask me for them before you start.
2. [[what to read, and what to note]]
3. [[how to score, e.g., PASS, FLAG, FAIL or UNKNOWN against each criterion, with evidence and a source]]
4. [[what to research, and where]]
5. [[what to rank, e.g., the top five risks, most serious first]]
6. [[what to ask, and whom]]
7. [[what to write at the end, e.g., a two-page memo with a recommendation]]

Output, always in this order: [[e.g., a one-line verdict; the scorecard as a table; the top risks; the questions; the memo]]. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.

Rules:
- [[e.g., Never state a fact without a source. Never invent numbers.]]
- [[e.g., Show the math for anything you calculate.]]
- You recommend; I decide.

Keep my criteria out of the skill: they stay in this project's instructions, so I can change them without rebuilding it. Build the skill from these instructions only; don't add scoring rules of your own.

If you can't save skills in this account, give me the same instructions as plain text that I can paste into this project's instructions.

Example: tonight's deal diligence agent

Create a skill called "deal-diligence" and save it to my skills.

Use it whenever I share an investment opportunity (a teaser, CIM, data room, term sheet or deal memo) and ask for diligence, a screen or a first look.

Steps:
1. Read my investment criteria from this project's instructions. If you can't find any, ask me for them before you start.
2. Read every document I share. Note the company, the sponsor, the terms, the financials and anything that's missing.
3. Score the deal against each criterion: PASS, FLAG, FAIL or UNKNOWN. For each, give one line of evidence and its source (document and section, or a URL). If the documents don't answer it, mark it UNKNOWN. Never guess.
4. Research the sponsor and the sector on the web: track record, litigation, news and typical valuation multiples. Cite every source. If you can't verify something, say so.
5. Rank the top five risks, most serious first. For each: why it matters and what would reduce it.
6. List the information gaps, then write fifteen questions for the sponsor, most important first.
7. Write a two-page memo: thesis, business, terms, returns, risks and a recommendation (Proceed, Proceed with conditions, or Pass), with the conditions.

Output, always in this order: a one-line verdict; the scorecard as a table (Criterion | Score | Evidence | Source); the top five risks; the information gaps; the questions; the memo. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.

Rules:
- Never state a fact without a source. Never invent numbers.
- When you calculate something (leverage, multiples, growth), show the math.
- Flag anything that relies on adjusted or pro forma numbers.
- You recommend; I decide.

Keep my criteria out of the skill: they stay in this project's instructions, so I can change them without rebuilding it. Build the skill from these instructions only; don't add scoring rules of your own.

If you can't save skills in this account, give me the same instructions as plain text that I can paste into this project's instructions.

Your own agent

[[AGENT NAME, e.g., DEAL DILIGENCE AGENT]]
When I share [[what you'll share, e.g., a teaser, CIM or term sheet]], do this:
1. Use the [[criteria, e.g., investment criteria]] above.
2. [[what to read, and what to note]]
3. [[how to score, e.g., PASS, FLAG, FAIL or UNKNOWN against each criterion, with evidence and a source]]
4. [[what to research, and where]]
5. [[what to rank, e.g., the top five risks, most serious first]]
6. [[what to ask, and whom]]
7. [[what to write at the end, e.g., a two-page memo with a recommendation]]
Output, always in this order: [[e.g., a one-line verdict; the scorecard as a table; the top risks; the questions; the memo]]. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.
Rules: [[e.g., never state a fact without a source; never invent numbers; show the math]]; you recommend, I decide.

Example: tonight's deal diligence agent

DEAL DILIGENCE AGENT
When I share an investment opportunity (a teaser, CIM, data room, term sheet or deal memo), do this:
1. Use the investment criteria above.
2. Read every document I share. Note the company, the sponsor, the terms, the financials and anything that's missing.
3. Score the deal against each criterion: PASS, FLAG, FAIL or UNKNOWN. For each, give one line of evidence and its source (document and section, or a URL). If the documents don't answer it, mark it UNKNOWN. Never guess.
4. Research the sponsor and the sector on the web: track record, litigation, news and typical valuation multiples. Cite every source. If you can't verify something, say so.
5. Rank the top five risks, most serious first. For each: why it matters and what would reduce it.
6. List the information gaps, then write fifteen questions for the sponsor, most important first.
7. Write a two-page memo: thesis, business, terms, returns, risks and a recommendation (Proceed, Proceed with conditions, or Pass), with the conditions.
Output, always in this order: a one-line verdict; the scorecard as a table (Criterion | Score | Evidence | Source); the top five risks; the information gaps; the questions; the memo. Put the whole result in one Markdown document (.md) that I can read in Claude right away, titled with the company name and today's date.
Rules: never state a fact without a source; never invent numbers; show the math for anything you calculate; flag anything that relies on adjusted or pro forma numbers; you recommend, I decide.

ChatGPT: paste it into your project's instructions, under the criteria.
Gemini: create a Gem (or a skill, if your Gemini has Skills) named deal-diligence. Paste this as its instructions, with your criteria at the top.
Copilot: Agent Builder › Describe: paste it, then add your criteria under Configure › Instructions.

Part B: Your criteria (save in the project's instructions, replacing the sample)

Your own criteria

My investment criteria:
1. Check size: [[$2M-$10M per deal]]
2. Sectors: [[sectors you invest in]]. Excluded: [[sectors you avoid]]
3. Geography: [[ ]]
4. Company size: [[EBITDA or revenue range]]
5. Revenue quality: [[e.g., at least 60% recurring]]
6. Profitability: [[e.g., reported EBITDA margin of at least 12%]]
7. Leverage: [[e.g., total debt at close, including any seller note, of no more than 4.0x EBITDA]]
8. Returns: [[e.g., 20% net IRR and 2.5x net MOIC with no multiple expansion]]
9. Hold period: [[ ]]
10. Sponsor track record: [[e.g., two completed exits we can verify]]
11. Sponsor commitment: [[e.g., at least 5% of the equity]]
12. Customers: [[e.g., no customer above 15% of revenue; 15-25% is a FLAG]]
13. Management: [[e.g., CEO committed for two years, or a named successor under contract]]
Deal-breakers: [[ ]]

Example: a fictional firm's criteria

My investment criteria:
1. Check size: $2M to $10M per deal.
2. Sectors: business and industrial services with recurring revenue; healthcare services. Excluded: consumer retail, biotech, crypto.
3. Geography: United States and Canada.
4. Company size: $3M to $25M of reported EBITDA.
5. Revenue quality: at least 60% recurring or contracted revenue.
6. Profitability: reported EBITDA margin of at least 12%.
7. Leverage: total debt at close, including any seller note, of no more than 4.0x EBITDA.
8. Returns: at least 20% net IRR and 2.5x net MOIC with an exit multiple no higher than the entry multiple. A base case that assumes a higher exit multiple is a FLAG; missing the thresholds at the entry multiple is a FAIL.
9. Hold period: 3 to 6 years.
10. Sponsor track record: at least two completed platform exits we can verify. Claimed but not yet verifiable is UNKNOWN.
11. Sponsor commitment: at least 5% of the equity.
12. Customers: no customer above 15% of revenue. 15% to 25% is a FLAG; above 25% is a deal-breaker.
13. Management: a CEO committed for at least two years after close, or a named successor under contract. A named successor without a signed contract is a FLAG.
Deal-breakers: any single customer above 25% of revenue; unresolved litigation not covered by insurance; no audited financials for the most recent full year.

Tip: where a criterion has a grey zone, say what counts as a FLAG, as the sample does. It keeps every run consistent.

Part C: Test it now (sample data)

Run deal diligence on this deal.

Files for the example company Project Firebreak, fictional

Deal packThe sponsor's summary: business, financials, terms. One document, three ways to get it: attach the PDF or the .md file to your chat, or copy the text and paste it in.

Download as PDForDownload as .mdor

Part D: Put it on a schedule (paid plans)

Your own schedule

Every [[Monday at 8am]], check the [[folder name]] folder in my [[Google Drive]] for deal documents added since the last run (teasers, CIMs, term sheets). For each new deal, run the deal-diligence skill using the investment criteria below, and give me the one-line verdict and the scorecard. Skip anything you've already reviewed. Don't reply to, share or send anything.

My investment criteria:
[[PASTE YOUR CRITERIA]]

Example: fictional folder and inputs

Every Monday at 8am, check the "Deal flow" folder in my Google Drive for deal documents added since the last run (teasers, CIMs, term sheets). For each new deal, run the deal-diligence skill using the investment criteria below, and give me the one-line verdict and the scorecard. Skip anything you've already reviewed. Don't reply to, share or send anything.

My investment criteria:
1. Check size: $2M to $10M per deal.
2. Sectors: business and industrial services with recurring revenue; healthcare services. Excluded: consumer retail, biotech, crypto.
3. Geography: United States and Canada.
4. Company size: $3M to $25M of reported EBITDA.
5. Revenue quality: at least 60% recurring or contracted revenue.
6. Profitability: reported EBITDA margin of at least 12%.
7. Leverage: total debt at close, including any seller note, of no more than 4.0x EBITDA.
8. Returns: at least 20% net IRR and 2.5x net MOIC with an exit multiple no higher than the entry multiple. A base case that assumes a higher exit multiple is a FLAG; missing the thresholds at the entry multiple is a FAIL.
9. Hold period: 3 to 6 years.
10. Sponsor track record: at least two completed platform exits we can verify. Claimed but not yet verifiable is UNKNOWN.
11. Sponsor commitment: at least 5% of the equity.
12. Customers: no customer above 15% of revenue. 15% to 25% is a FLAG; above 25% is a deal-breaker.
13. Management: a CEO committed for at least two years after close, or a named successor under contract. A named successor without a signed contract is a FLAG.
Deal-breakers: any single customer above 25% of revenue; unresolved litigation not covered by insurance; no audited financials for the most recent full year.

All files for the example company Project Firebreak, fictional

Deal packThe sponsor's summary: business, financials, terms. One document, three ways to get it: attach the PDF or the .md file to your chat, or copy the text and paste it in.

Download as PDForDownload as .mdor
Investment criteriaThe firm's 13 criteria and deal-breakers

Download .mdor
Sponsor modelThe sponsor's Excel model for the deal

Download .xlsx
Options positionsStock, cash-secured puts and a covered call

Download .csvor
Options chainPuts and calls to pick new strikes from

Download .csvor

Build your first agent tonight

Get Claude ready

Design it before you build it

Give it memory: your criteria

Write the job description

Run it on a real-looking deal

Make it yours

Let it run without you

Three more agents, one paste each

Model analysis

Financial modeling

Options trading

Deal diligence

From your AI to your company's AI

Book a complimentary setup call