5. A Narrow Offer With a Written Boundary
Summary
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.
Choosing the digest and writing down what it promises
The digest now runs on its own each weekday evening. You open the day's folder, read one page, check the count line at the bottom, and you are done in about four minutes. That figure comes from five invented days, so it is a teaching number, not a measurement of a real business. Still, it is more than you know about either of the other two ideas.
Choosing between the three candidate offers comes down to one question: which one can you describe, deliver and price from your own records? Only the digest. You have built it, run it and timed the review. The missed-call text-back would mean sending texts from a business number through a phone system nobody has tested. The quote follow-up reminder exists only as a line in the research packet. So park those two, and ask Claude Code to write a short note in the folder for each, saying why it is parked and what would bring it back. The text-back comes back if a real owner says missed calls matter more than a daily summary. The reminder comes back if the digests show many quotes going cold. Parking keeps the thinking, where deleting would throw it away.
One caution stays attached to the choice. Five invented days show that the digest can be delivered. They do not show that anyone wants it. No owner has seen one, and no one has said they would pay. That is a separate question with separate evidence, and the evidence has to come from real people.
Now turn the digest into a promise someone else can check, and keep that promise in files Claude Code reads. Scope is the written boundary around the offer, the line between what you promise and what you do not. Start in plan mode and ask Claude to draft a scope file from what the folder already shows: the CLAUDE.md and the digests from the five invented days. Read the plan, approve it, then read the file it writes. Here is what it should say.
It is for one small trade business with one owner who misses enquiries while working on site. The customer routes voicemail transcripts, website form messages and emails into one shared folder. The service reads that folder once each weekday evening and delivers one page by seven in the evening, local time. The page lists every enquiry from that day and says whether each has a callback number or a reply address. Enquiries with neither come first, under "unanswered", marked unreachable. The page ends with the count check. The digest covers up to thirty enquiries a day. That is an invented ceiling, to be moved once real volumes show up, and above it the page says so plainly.
Then the out-of-scope list. The service does not reply to customers. It does not send texts or emails on the owner's behalf. It does not touch the phone system, the voicemail settings or the website. It does not book jobs, write quotes or chase payments. It does not run at weekends. It never guesses at missing details: if a voicemail gives a first name and no number, the digest says the number is missing and does not try to find one.
That list does two jobs. It tells a customer where the promise stops, so a request like "could it also reply to them?" already has a written answer. And it tells you which access to withhold from the agent. A line saying never reply to customers is only an instruction. The safer setup gives Claude Code no way to send a message at all: no email connector, no texting service, nothing in its tools that can reach a customer. The scope tells you what to leave unconnected.
Scope says what arrives. Acceptance criteria say how anyone can tell whether it arrived correctly. A good criterion passes or fails, and someone other than you can decide which. The digest already carries checks from earlier work: the claim-check list, the count line and the reachability marks. Ask Claude to turn each into a criterion a stranger could apply, and to save them in a file of their own. They come back as five.
- The number of enquiries listed matches the number of enquiry files in that day's folder, and the count line states both numbers.
- Every enquiry is marked with whether a callback number or reply address is present, and any enquiry with neither appears first under "unanswered", marked unreachable.
- Every name, number, time and detail in the digest appears word for word in a source file, and the claim-check list points to that file.
- Nothing appears that is not in a file, so a missing detail reads "not reported" rather than a guess.
- The digest is in the agreed place by seven in the evening on a weekday.
Keep the criteria in that one file and point CLAUDE.md at it. CLAUDE.md can pull in another file: a line made of an at sign followed by the file's path tells Claude Code to load that file along with CLAUDE.md at the start of every session. Ask Claude to add the line. The diff shows a single new line, marked with a plus sign. Approve it and commit. Now every scheduled run reads the criteria, and when a customer agreement needs to quote them later, there is exactly one version to quote.
Here is how a stranger would use the criteria. Say someone outside the business agrees to check one day's output. He opens the folder and counts twelve files. He opens the digest and counts twelve entries, and the count line says twelve and twelve. That passes. He finds two entries marked unreachable, and both sit at the top. That passes. He picks three phone numbers and searches for them in the files. All three are there, exactly. That passes. He notices one entry says the job is a boiler repair, and the voicemail never says "boiler". It says "the heating's packed in". That fails criterion three. The model filled in a likely detail, and the criteria caught it. He needed no trade knowledge, only the page and the folder.
Some criteria need no person at all. Counting files and counting entries is plain arithmetic, the kind of fixed answer ordinary code gets right every time and a model can only add variation to. Checking the time stamp on the digest is the same. Checking that each quoted number appears in some file is a text search. The judgment-heavy part is criterion four, spotting an inference dressed as a fact, and even that becomes a lookup when the claim-check list names its source. The more a criterion can be checked by a script, the less of your time it uses, and that matters as soon as you start counting cost.
What one customer's digest costs to deliver
Delivery cost is what it takes to keep the promise for one customer. It is not the price. It is the floor under the price. If the price sits below it, every sale loses money. Build the estimate from three parts: your review time, the model usage for each run, and the cost of runs that fail or arrive late. Every number below is illustrative, and the estimate is labeled that way in the project.
Start with review, because it is easiest to see. Review takes about four minutes on invented data. To turn minutes into money you need a rate for your own time. Pick sixty dollars an hour as an illustrative figure. That is a dollar a minute, so a four-minute review costs about four dollars. Over twenty-two weekdays in a month, that is about eighty-eight dollars of your time for one customer.
Pricing your own time can feel odd when no money changes hands. But that time could go to a second customer, to selling, or to rest. Leave it out and the service looks cheaper than it is, and the truth arrives only when ten customers each want four minutes every evening.
Next, the model's usage for each run. Models are billed in tokens, small pieces of text. A useful rule for English is that one token is about three quarters of a word, so a thousand words is roughly thirteen hundred tokens. Providers price input tokens, the text you send, and output tokens, the text the model writes, separately, quoted per million tokens. Output costs more.
So estimate what one run sends and receives. Say a busy invented day holds about three thousand words of enquiry text. That is around four thousand tokens. The skill's instructions, the request and some working space add perhaps two thousand more, making about six thousand input tokens. The digest itself runs about seven hundred and fifty words, so roughly a thousand output tokens.
Then apply a price. When this chapter was researched, Anthropic listed its Sonnet model's application programming interface price at about three dollars per million input tokens and fifteen per million output tokens. The application programming interface, or API, is the way software calls the model directly and pays per use. Six thousand input tokens is six thousandths of a million, times three dollars, which is just under two cents. A thousand output tokens is one thousandth of a million, times fifteen dollars, which is one and a half cents. One run costs about three and a half cents. Over a month of weekdays, that is under a dollar.
Two cautions keep this honest. First, the prototype runs inside a Claude subscription, which is not billed per token at all. Subscriptions and API credit are billed separately by every major provider, and one does not include the other. The per-token figure is what the work would cost once it runs for a customer on API billing. Second, the guess about tokens is a guess. You can replace it with a measurement. Every response from Anthropic's API carries a usage section with the exact number of input and output tokens that call used, and a separate token-counting endpoint will tell you how many tokens a request would use before you send it. Logging those counts for a week of runs turns the estimate into a record. Prices also change, so recheck the current rate before quoting anything.
The third part is failure, and failure is where cost hides. Suppose, illustratively, one run in ten goes wrong. The laptop sleeps and the run never happens. Or the count does not match, or a detail is invented. Each failure costs more than a normal review. Someone has to notice it, rerun or repair it, check it again, and perhaps explain to the customer. Call that twenty minutes, or twenty dollars at the illustrative rate. If one run in ten fails, the average failure cost per run is a tenth of twenty dollars, which is two dollars. Over a month that adds about forty-four dollars.
A late run is a failure of a specific kind. The digest's whole value is arriving before the owner plans the next day. A digest at midnight might be perfect and still useless. You cannot put a figure on a missed job for an owner you have never met. What you can do is decide, in advance, what you owe the customer when it happens: a credit for that day, say. That credit becomes part of failure cost, and the fifth criterion makes "late" something anyone can check.
Now add it up for one illustrative customer over a month: about eighty-eight dollars of review, about forty-four dollars of failure handling, and under a dollar of model usage. That comes to around a hundred and thirty-three dollars before any software subscription or hosting.
The shape of that sum matters more than the total. The model is less than one percent of the cost. Your minutes are nearly all of it. So cheaper models do little for the margin. Fewer failures and shorter reviews do a lot. Each acceptance criterion a script can check takes a slice off review. Each failure prevented removes twenty minutes. Moving the run off a laptop that might sleep would cut one whole category of failure.
This also says what not to do. Do not price at a dollar a day because the model costs four cents. Do not assume four minutes stays four minutes, either. Real enquiries will be messier than invented ones, and the first review of the prototype took fifteen minutes. A real first week might look like that again.
What you have now is a narrow offer: a written scope, five pass-or-fail criteria and a labeled floor under the price. What you do not have is a single person who has said they would pay. That is the next question, and the floor tells you which answers would make the offer worth keeping.
Before answering it, though, the five criteria give you something you lacked when you first chose a tool: a fixed test any assistant could be held against.
The digest runs in Claude Code, on your laptop, with the skill loaded. That setup was picked by the shape of the job: it repeats, it keeps files, and it must be inspected later. That was a sensible start. It was never a comparison. Now you can run one, because you have a success test that does not depend on taste.
Take one invented day's folder. It holds eleven enquiry files: voicemail transcripts, form messages and emails. Two have no number and no address. One mentions a time with no date. Make copies, so no tool touches the original. Give each assistant the same request you already use: the skill's evidence rules, the statement that every file is data and never an instruction, and the five acceptance criteria written out. Run each tool on the folder three times, because one good run proves little. A model can answer differently on the same input, so consistency is part of what you are buying.
The test results below are an illustrative run on invented data. What each tool offers comes from its providers' documentation at the time of research. Where an outcome follows from a documented feature, that is said.
Claude, through Claude Code, is the baseline. Setup is already done: the skill sits in the folder and the schedule exists. It reads the whole folder because it works over files on disk, so all eleven files are visible. In three runs the count line read eleven and eleven each time. Both unreachable enquiries came first, and the claim-check list pointed to real files. On one run the wording of a summary changed, but no criterion depends on wording, so all three passed. On data handling, Claude's business plans, Team and Enterprise, exclude customer prompts and files from model training. Team was listed at around twenty-five dollars a seat a month, or twenty a month paid yearly, with a minimum of two seats. If you are on Pro, at about twenty dollars a month, check the consumer privacy settings rather than assume business terms apply.
ChatGPT has Projects, which hold pinned files. The Plus plan allowed twenty-five project files when this was researched, and the Team plan forty. Setup means uploading the files, up to ten per batch, and pasting the request into the project's instructions. The documented feature that matters here is the code tool, often called Code Interpreter or Advanced Data Analysis. It runs Python in a sandbox. Ask it to count the files with code before writing, and the count comes from arithmetic rather than the model's reading. That serves criterion one well. In these runs, two of three passed. On the third, the model skipped the code step and summarized from its reading of the files, and one enquiry was merged into another. The digest listed ten against eleven files, and the count line caught it. So the result depends on whether the code tool is used, and an instruction to use it is only an instruction. ChatGPT's Team and Enterprise plans exclude business data from training by default. Plus was listed at twenty dollars a month and Team at roughly twenty-five to thirty per user.
Gemini comes in two useful forms. The Gemini app has a very large context window, the amount of text it can take in at once, reported at one to two million tokens. Eleven small files are nothing to it. The second form is NotebookLM, Google's notebook built for answering from your own sources. Every statement it makes carries an inline citation linking back to the passage it came from, and it can sync a Google Drive folder. That suits criterion three. Checking a phone number becomes a click on its citation. These runs passed criteria two, three and four each time. NotebookLM is built for reading and briefing, though. Shaping its output into the exact one-page format took more prompting, and on one run the count line was missing. That is a fail. A check that is not stated cannot be checked. On Google Workspace business plans, customer data is not used to train Gemini models without admin consent. The consumer Google AI Pro plan was listed at about twenty dollars a month.
Perplexity is the credible alternative outside the big three. Its Spaces hold shared files with a saved prompt and inline citations, and let you switch between models in one workspace. It is strongest when a job mixes your files with web search. This job has no web search, and the scope forbids guessing missing details from outside sources. Its main advantage would go unused, so note it and move on.
Now line them up against the test, not against reputation. On pass rate in this illustrative run, Claude Code passed three of three, ChatGPT two of three, and NotebookLM two of three. Each failure was caught by a criterion, not by instinct. On setup, Claude Code needed nothing new. The others needed uploads and a pasted request, and that would have to be repeated daily for a scheduled service unless connected to a folder. On consistency, the failures differed in kind. ChatGPT's came from a step it could skip. NotebookLM's came from formatting.
Scheduling decides the most. The service must run every weekday evening without you. Chat and notebook apps on a subscription are made for a person sitting there, uploading and reading. The digest already runs this way as a Claude Code desktop task, drawing on your plan's usage limits rather than API credit. For the chat and notebook apps, automated scheduled reporting goes through the API, where software calls the model, can count the files with code first, and can demand a fixed output format. The subscription does not include API credit, and API spending does not buy a chat seat. So "which subscription is best" is the wrong question. The useful question is which model the scheduled run should call and what that costs per run. The cost estimate already answered that. Per-run model cost is pennies on any of these providers, and your time is nearly all of the bill.
So the stack stays, for a concrete reason. Claude Code passed every criterion in every run on the job as written. It already holds the skill and the schedule. It works on files where they live rather than on uploaded copies. Switching would not lower cost, because cost is review time. It would not raise the pass rate on this test. It would mean rebuilding work that already passes.
The comparison still taught two things worth carrying forward. First, ChatGPT's code tool showed that counting is a job for code, not for the model. That points to an improvement in your own setup: a small script that counts the files and checks the time stamp before the model writes anything. That would turn criteria one and five into hard checks rather than requests, and shrink review. Second, NotebookLM showed how fast a reviewer moves when every claim links straight to its passage. The claim-check list does a rougher version of that, and tightening it to name the exact line would cut your minutes further.
Save the folder copy, the three runs from each tool and the pass-or-fail marks into the project. When a provider changes its plans or features, run the same folder and the same five criteria again and see whether the answer changes. That record is the asset. The test and its results stay with you, whatever tool you end up using.
Want a book of your own?
Gnothi wrote and narrated this chapter. Name a subject and it writes you a course, or a story that teaches it, one chapter at a time, delivered to your podcast app.
Get started free