Agentic AI Course

With one invented trial customer now paying, the chapter builds three pieces to find the next: a claim-checked offer page published on GitHub Pages, a captioned sample demo, and a HubSpot record of prospects and follow-ups. Each is verified where it lives, and none has yet met a real owner. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Closing out the trial and building the offer page In the example, one invented customer is paying for a trial, which does not say where the second customer comes from. The chapter builds three things for that: an offer page, a short demo, and a record of who to contact next. The customer, runs, prospects and money are all still the show's invented example, and nothing here proves real demand. Before building, the example's trial week is checked. The completed-runs record should hold five keys, one per scheduled date with no repeats. The customer record's delivered tally should be up by five, with the late-day credit field still empty. The two files are written by different steps, so agreement is evidence. They agree, the invented heating engineer says yes in writing, and the reply is logged with its date and how it arrived. The page is drafted in Claude Code from the scope file alone, with one audience (single-owner trade businesses), one offer (the end-of-day digest) and one action (book a ten-minute look at a sample). The claim-check flags a line about winning back jobs as unsupported, so it is replaced with a plain description. No testimonial is used, since none exists. Checks by hand cover contrast ratios under WCAG 2.2 (the WCAG 2.2 quick reference), heading order, image descriptions, and tabbing through the page. Lighthouse's accessibility score covers only machine-checkable audits, and WAVE cannot judge description quality, so neither replaces the manual checks. Publishing with GitHub Pages The page lives in its own small repository and is published from a branch. Pages sites are public, so the customer record, payments and notes stay out of that repository. Pushes made with a workflow's built-in token do not trigger a build. A custom domain is entered in Pages settings first, then DNS is set, and propagation can take up to a day. A captioned demo The demo is about a minute, narrated from a script and recorded in Clipchamp using the labeled sample day, with "invented sample data" on screen throughout. Automatic captions need a voice track and misfire on key words, such as "digest" becoming "diet chest", so each line is corrected by hand. Prerecorded captions are a basic accessibility requirement. Whether to burn captions in or ship a separate file is left to check against current guidance. Prospects in HubSpot HubSpot's free CRM holds three invented prospects (a plumber, a roofer, an electrician), each with a company and a dated follow-up task, through the HubSpot connector for Claude. It needs a paid Claude plan and admin approval in HubSpot. HubSpot also offers a separate remote MCP server. Write permissions and Claude Code setup should be checked in current documentation. The agent's report is not trusted: each record is opened in HubSpot. In the example, one task had no due date and is fixed and rechecked. None of the three pieces has met a real owner yet.

The enquiry digest moves into a cloud scheduled task so it runs with the laptop closed. The chapter covers alerts for failed or late runs, a completed-runs record that stops duplicate sends, and the monthly offer to make when the trial ends, using invented example figures. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Running the digest with the laptop shut The enquiry digest only ran when the laptop was awake and the Claude Code app open, and a failed run told no one. The customer, dates and money remain the show's invented example and prove nothing about real demand. The first fix is a cloud scheduled task, formerly called a routine and still a research preview, so check current documentation. The setup covers a self-contained prompt, a schedule, a GitHub repository and an environment. Cron times are read in UTC unless you add a time zone, and the shortest interval is an hour. Schedule a few minutes off the hour, well before the seven o'clock deadline. The Google Drive server added on the laptop is invisible to the cloud run. Add it as a claude.ai connector or commit a .mcp.json file. Keep API keys out of environment variables by using network secrets and the environment's network access levels. A cloud run uses auto mode, so the read-only Drive scope and the Viewer share are the limits that still hold. Whether the laptop's prompt hook carries over is unconfirmed. Detecting failed runs and avoiding duplicate sends A Succeeded label in History only means no failure was reported. For headless runs, check the exit code, printed output, and MCP errors. Also check the result subtype and stop reason. The wrapper script that combines them is the show's own pattern, not an Anthropic script, so test each condition. Stop and StopFailure hooks can add a quality gate and alert logging. A separate check shortly after seven catches a run that never started. For retries, each job gets a key from the customer's opaque identifier and the scheduled date, recorded in a small file in the repository. The digest is written to a fixed path, so a rewrite replaces the earlier copy. One gap remains: a crash after delivery but before the key is marked done. The owner's review before sending covers it during the trial. Stripe's idempotency keys solve the same problem, but keys may be discarded after a day. A GitHub Actions scheduled workflow can run the plain file checks, though it can delay or skip runs and does not prevent duplicates. What to offer when the trial ends The trial cost $135 for ten digests, so each delivered digest earns $13.50. With four delivered, $54 is earned and $81 is still owed in digests. The delivered count in the customer record should match the completed-runs record. Claude Code drafts the offer from the scope file and price record only: an illustrative $270 a month, just over twice the illustrative floor of about $133. The trial agreement said it would not roll into a monthly charge, so billing starts only after written agreement. Without a yes, the trial ends. A discount toward $100 would earn less per digest than it costs. Before sending, the cloud task runs a clean week under watch.

A worked example of taking a first paid trial for the enquiry digest: checking and recording the Stripe payment, connecting the customer's Drive folder read-only, and adding a hook script that checks each run before the model writes. It ends by tracing the payout into the bank and separating cash received from revenue earned. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Taking the first payment The chapter's customer is invented: the heating engineer from the interview record, with made-up money, dates and enquiries. A pretend payment says nothing about real demand. It is used to show how the process works. The chapter only goes ahead because the owner made a written, dated change to the commitment rule after the answers were in. Without that change the digest is parked. Before sending the $135 two-week trial request, the owner checks three things personally: the recipient, the amount, and the bank account the payouts go to. The engineer pays. Stripe shows the payment as Succeeded, with a gross of $135, a fee of about $4.22 and a net of about $130.78 for a domestic card. Claude Code adds the first row to the payments record. It confirms that gross minus fee equals net and leaves the payout cell blank. The chapter explains Stripe's chain of identifiers: payment intent, charge, balance transaction, payout. Onboarding uses a Google Drive folder the customer shares as Viewer. It is reached through the existing read-only MCP server, using a test account rather than a personal one. Digests go to a delivery folder in the project and are reviewed before they are sent. A gate before the model runs A script handles the steps that have one right answer: whether the folder can be read, how many files arrived that day, and whether the run is late. It runs as a UserPromptSubmit hook, because a hook that runs at session start cannot block the session. The script's exit code controls the outcome, as described in the Claude Code hooks reference. Exit 0 lets the run go ahead. Exit 2 stops it with a stated reason. If the run is late it still goes ahead, with a note that the late-day credit applies. Because the hook fires on every prompt in the folder, the script first reads the prompt from the JSON it receives and checks only the digest run. Its one dependency, Google's Python library for the Drive API, is installed with pip and listed in a requirements file. The run's permissions are settled in advance and limited to Drive list and read plus writes to the delivery folder. The chapter separates three things: the model, Claude Code acting as the agent runtime, and the system of record, which is the payments file and a new customer file. The first invented Monday brings seven files and the digest passes all five criteria. Tracing the cash A new Stripe account holds its first payout for about one to two weeks. After that, card funds usually become available in two business days, as the payout schedules FAQ describes. The bank deposit shows only a lump sum. The itemized payout reconciliation report links each payout to its transactions, so Claude Code can fill in the payout identifier. Any mismatch is brought to the owner along with the evidence. Cash received is kept separate from what has been earned: each of the ten digests earns a tenth of the fee. Three things remain open: Desktop tasks skip their run when the laptop is asleep. Cloud scheduled tasks avoid this. A failed run doesn't alert anyone. The trial's end date is coming, and the owner needs to decide what to offer before it arrives.

Claude Code turns invented interview notes into a checked record, writes a first Python script that confirms every quote against its notes file, tests dated stopping rules under git, drafts the trial terms and builds a $135 Stripe Payment Link in test mode with a restricted key. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Filling the interview record with Claude Code Every owner and answer in this chapter is invented to show the method. Rough notes go into one plain text file per owner, committed to git. In plan mode, with the evidence-checked-drafting skill called by name, Claude Code fills a comma-separated values record with one answer per row: owner, question, exact quote, date, how it was given, commitment and the notes file it came from. A quote that contains a comma sits inside double quotation marks so a spreadsheet keeps it in one column. In the invented spread of five owners, only a heating engineer agrees to the paid trial. Two owners show polite interest and two commit nothing. A first script that checks the quotes The skill's claim-check list is the model reporting on its own work, so Claude Code writes a short Python script that reopens each notes file and confirms every quote appears in it exactly. It needs only Python and its built-in csv module. The chapter reads the script as a variable, a loop over rows, indentation, a condition and a printed total. The first run catches a quote the model tidied from "Would've" to "Would have". A false failure from a phone's curly apostrophe is fixed by a small change reviewed as a diff, and a Python traceback is read from its last line up. The agent's summary claim of "strong interest from three owners" is cut back to one commitment and four owners who committed nothing. Stopping rules under git Claude Code tests the three stopping rules against the record by row, and git log shows the rules file was committed before the first notes file. The price rule and the text-back rule do not fire. The commitment rule, read as written, says park the digest. Going ahead needs the rule reworded to "if no owner commits, park the digest" and committed with a message saying it changed after the answers, so git show displays the old and new wording side by side. Trial terms and a payment link built in test mode The one-page agreement is drafted only from the scope file. The owner decides three gaps it flags: a $135 trial price, payment before the first digest, and no automatic roll into a monthly charge. Clauses that depend on jurisdiction are marked for an authoritative local source or a lawyer. Claude Code then builds the product, price and Payment Link with the Stripe CLI in test mode, using a restricted key that can create only those three things. Stripe counts dollars in cents, so the price command reads 13500. The owner pays the test link with Stripe's test card, then creates the live link in the Dashboard and leaves it unsent. At 2.9% plus 30¢ per domestic card charge, about $130.78 of $135 reaches the balance, and refunds do not return the fee. A new account's first payout waits seven to fourteen days, and after that payouts follow a schedule of about two business days in the US. Each payment carries a PaymentIntent identifier, and the new payments record has its column headings and no rows yet.

The chapter turns an invented cost floor into a testable first price, builds a self-checked sample, and has Claude Code write the interview files and dated stopping rules before any owner is asked. It then connects Claude Code to a Google Drive folder through an MCP server with three separate read-only limits and verifies what the agent actually did. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website A first price and a checked sample Every figure here is invented to show the reasoning, not to report what real trade owners pay. The illustrative delivery-cost floor of about $133 per customer per month leaves out card fees and onboarding, so it runs a little low. The illustrative first price is $270 a month, just over twice the floor, a gross margin of about fifty percent to pay for finding customers, creeping requests and hard weeks. Revenue, profit and owner time are tracked as three separate numbers. Using the evidence-checked-drafting skill, Claude Code drafts an offer paragraph only from the scope file, and the claim check deletes an unsupported line, "saves owners an hour a day." A sample digest from one invented day passes all five criteria. Files written before anyone answers Claude Code drafts an interview sheet with three questions: what missed enquiries cost the owner now, whether they would start a paid two-week trial at the stated price, and what would have to change for a yes. A comma-separated values record gives each answer its own row with an exact quote and a commitment field, where praise is logged as polite interest. Three stopping rules go into their own file and a git commit before any interview, so the commit's date shows they came first. A read-only Google Drive connection A real customer's enquiries would land in a shared cloud folder. Claude Code reaches Google Drive through an MCP server rather than a browser agent. Setup in the Google Cloud console takes a project, the Drive API switched on, an OAuth consent screen with the test account as a test user, and a desktop-app OAuth client whose credentials file stays outside the repository, following setups like the Google Drive MCP server setup for Claude Code. The server is registered with claude mcp add, checked with claude mcp list, and its tools are inspected in the session's mcp panel. Three separate limits keep it read-only: the drive.readonly OAuth scope, Viewer access on the folder, and a tool list with only list and read. The agent lists nine files, the tool calls and Drive agree, and a rename is impossible.

The chapter chooses the end-of-day enquiry digest, has Claude Code write its scope and five checkable acceptance criteria into files that CLAUDE.md imports, and works out an illustrative cost floor dominated by review time. It then tests Claude Code, ChatGPT and NotebookLM against the same criteria, sets Perplexity aside, and keeps the current setup for concrete reasons. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Picking the digest and writing its promise into files Of three candidate offers, the end-of-day enquiry digest is the only one that can be described, delivered and priced from the folder's own records. The missed-call text-back and the quote follow-up reminder are parked, each with a note on what would bring it back. Five invented days show the digest can be delivered, not that anyone wants it. In plan mode, Claude Code drafts a scope file covering one trade business, three kinds of enquiry in one shared folder, one page by seven each weekday evening with unreachable enquiries first, and an invented ceiling of thirty enquiries a day. The out-of-scope list rules out replying, texting, booking, weekend runs and guessing missing details, and it names which access to withhold from the agent, because a written rule is not a limit. Five pass-or-fail acceptance criteria go into their own file, and CLAUDE.md imports that file with an at-sign line so every run reads one version. A stranger comparing the page with the folder catches a "boiler repair" that a voicemail never mentioned. The floor under the price With illustrative figures, four minutes of review a day at sixty dollars an hour comes to about eighty-eight dollars a month. Failure handling adds about forty-four. Model usage at Sonnet API rates of roughly three dollars per million input tokens and fifteen per million output tokens costs under a dollar. The token estimate assumes about three quarters of a word per token, following OpenAI's guide to what tokens are and how to count them. Each API response reports its exact input and output token counts, so a week of logged runs can replace the estimate. A subscription does not include API credit. The model is under one percent of the cost, so shorter reviews and fewer failures matter far more than a cheaper model. Holding the assistants to the same test The chapter runs one eleven-file folder three times through each of three tools, using invented data. Claude Code passed all three runs. It already holds the skill and the schedule, and on Team plans it excludes business data from training (Claude pricing for individuals and organizations). ChatGPT Projects passed two of three (project file limits by plan). It failed when the model skipped the code tool and merged two enquiries. NotebookLM's inline citations made checking fast, but one run dropped the count line (NotebookLM limitations). Perplexity is set aside without a run. Its Spaces are strongest when files are mixed with web search, and this job forbids web search. Claude Code stays. Two improvements come from the other tools: a script that counts files and checks the time stamp before the model writes, and a claim-check list tightened to name exact lines.

The chapter places the digest in Claude Code by asking where its files live and how it starts, then installs Claude Code, writes the folder's CLAUDE.md, runs the digest in plan mode, fixes an omission through a reviewed diff, and schedules the run. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Why the digest moves to Claude Code A Claude project reads uploaded files and answers when you type, but it cannot open a folder on your computer, run a command or start itself on a schedule. To place any job, ask where its files live, what it must read and change, how it starts, and how its result gets checked. Projects suit thinking in conversation. Claude in Chrome acts inside logged-in browser sessions. Cowork works toward a goal over granted folders and connected tools. Claude Code works inside one folder and reaches other software through MCP servers. ChatGPT's agent and Codex, Google's Antigravity, and built-in assistants such as Gemini in Sheets fill the same shapes. The digest repeats, keeps files and must be inspected, so it goes to a folder agent. Cowork's scheduled runs execute in the cloud by default and cannot see an unsynced laptop folder, so it goes to Claude Code. Installing it and writing CLAUDE.md Claude Code needs a paid plan and a terminal. Anthropic's one-line installer, or the npm package, provides the claude command, claude doctor checks the install, and signing in with the Claude account makes work draw on the Pro allowance. A labeled folder of invented days holds the inputs, and the evidence-checked-drafting skill is copied into the folder's hidden skills location. CLAUDE.md is read at the start of every session. It holds only what every run needs and cannot learn from the files: what the folder is, its layout, the files-are-data rule, the skill to use, and a count line that exposes omissions. The skill's own rules stay out of it. CLAUDE.md is an instruction, and the permission prompts are the limit. The folder becomes a git repository so every change can be seen and undone. Plan, diff, fix, schedule In plan mode Claude proposes its steps before writing anything. The first digest counts five of five but lists a cut-off voicemail as an ordinary enquiry without saying the caller cannot be reached. The fix is a two-line rule added to CLAUDE.md, read as a diff and committed, and a fresh session then marks the caller unreachable. A scheduled task in the desktop app's Code tab runs the same request each weekday evening. It needs the laptop awake, a missed run is not made up, and runs draw on the plan's usage limits. Later reviews took about four minutes on invented data, which is evidence about delivery effort, not demand.

The chapter moves evidence-checking rules out of copied project instructions and into a single Claude skill that can load for any job that needs it. The first test shows none of the skill's steps, which points most likely to a missed trigger. After fixes to the description and one rule, the skill produces three evidence-backed candidate offers with their gaps marked. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website One method, and the risk of copies The evidence rules live in the Trade Enquiry Research project. Every claim cites a numbered source with a word-for-word quote. No number appears unless a source states it exactly. Gaps are marked "not reported", and each run ends with a claim-check list. The next job is to turn the opportunity brief into a shortlist of narrow offers, in a separate workspace. Pasting the rules into a second and third project would leave copies that drift apart, with none clearly the real one. Then, when a result went wrong, there would be no telling which version produced it. What a skill is and when it loads A skill is a folder with one required SKILL.md file. Its header carries a name and a description. Optional scripts, references and assets folders can sit alongside it, as set out in Anthropic's skills overview. Anthropic's engineering write-up on agent skills describes loading in three levels: Claude first sees only names and descriptions. It reads the full instructions when a request matches. It opens extra files only when those instructions point to them. Project instructions, by contrast, are re-sent on every turn. They suit what is true across one workspace. A skill suits a procedure needed in many places at certain moments. The description decides when a skill loads, so it should say what the skill does, what should trigger it, and what it is not for. The chapter also covers plan availability, the code execution setting, and uploading a zipped folder, following Anthropic's guide to creating custom skills. A likely missed trigger, then a fix The first test returns fluent offers with no claim-check list. One offer claims owners "win back jobs they would have lost" with no source behind it. Every step was missing, so either the skill never loaded or the model ignored all of it. A missed trigger is the likelier cause, so the chapter tests that first. The description spoke of briefs and recommendations, while the request asked for offers. The rewritten description covers options, shortlists and existing briefs. A new rule says any stated benefit needs its own quote or becomes an open question, which also covers a skill followed loosely. A negative test stays quiet. The repeated request returns three offers, each labeled illustrative and untested: a missed-call text-back a same-day quote follow-up an end-of-day enquiry digest Each comes with quotes and questions marked "not reported". Nothing says any owner would pay. The duplicated rules are then retired from the research project, and a check confirms the skill still runs there.

How a Claude project builds what it sends each turn, what the instructions field should hold, what memory does and does not carry between chats, the two-hundred-thousand-token window and the switch to retrieval, with short notes on ChatGPT Projects, Gemini Gems and NotebookLM. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Instructions, memory and filing Claude's project instructions are pasted in front of every conversation in the project on every turn, so they are re-sent rather than remembered and cost room each time. A worked example with invented figures shows a two-thousand-token block as one percent of the window per turn. The field should hold only the role, the output structure and the evidence rules, and the chapter reads out a trimmed version for the research project. Each chat starts from the files and instructions alone. With memory and past-chat search turned on, Claude can also recall and look up other chats in the same project, but that is recall, not a filing system. The working loop stays manual: produce, check, save locally, upload. Files may be up to thirty megabytes in plain text, Markdown, PDF, Word or CSV, and PDFs over a hundred pages are read as text only. Fixed note headings with a "not reported" rule make sources comparable, prefixes keep synthetic files out of findings, and questions with one fixed answer, such as counting files, belong to a script. What the model actually sees Text is counted in tokens. On paid plans the context window is about two hundred thousand tokens, roughly five hundred pages, and it holds the instructions, the conversation, the files and the answer. While the project fits, every file is present in full. Past that, paid plans switch to retrieval, which searches the files and hands the model relevant chunks, with up to ten times more capacity but no guarantee a file is read. That argues for short single-subject files named for their subject. A file limit is enforced by the software; "answer only from these files" is an instruction the model usually follows. Uploaded text is data, never authority. Other workspaces ChatGPT Projects lets chats in a project reference each other through project-only memory and allows five-hundred-and-twelve-megabyte files. Gemini Gems take up to ten files, and Google documents a million-token window for them. NotebookLM has no instructions field but answers only from its sources. The brief, headings and filing loop carry over to each.

The chapter sets up a Claude project with a chosen plan and model, two numbered source files, a five-part brief and saved instructions, then catches an invented statistic and reruns to an honest brief that ends in six open questions. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Claude, a project and the model picker Claude is Anthropic's assistant, used here in its web and desktop app. The chapter sets out what to pick before starting: a plan (the course assumes Pro, listed at twenty dollars a month in autumn 2026) and a model (Opus, Sonnet or Haiku, with Sonnet for this job), plus what the extended-thinking and web-search switches do. It then creates a project called Trade Enquiry Research, uploads two numbered source documents about an invented productized service for trade businesses that miss calls while on site, and explains context as the material in front of the model on one run. A five-part brief and a caught invention The brief names the deliverable and its five sections, limits evidence to the uploaded files, requires a quoted source for every claim, turns web search off, and asks for a claim-by-claim check at the end. The chapter reads the whole brief out as the message to send. The first result slips in a "forty percent" figure that no source contains and smooths over eight sources that say nothing about money. Instead of arguing with the output, the chapter writes five standing project instructions that apply to every future chat. The rerun is half the length and ends in six questions only a real conversation can answer. Model, application and agent The trained model only continues text; the app decides what the model sees; a tool-using agent acts in other software. Nothing in this chapter was an agent. To check a claim, open the uploaded file in the project and compare the quote yourself.