The prompt is the least important part of a Claude workflow. The exports, the rules file and the checklist it runs against are what make the same task come out the same way in October as it did in March.
There are a lot of "50 ChatGPT prompts for Amazon sellers" posts. This is not one, because prompts alone did not work for us. What worked was building a small amount of infrastructure once, a Project per brand with the rules and product facts loaded, a handful of skills for the tasks we repeat, and connectors to the tools that hold the data, and then running the same fifteen tasks through it every week. This post is that list, with enough detail to copy.
Evolve Media Agency sells AI search visibility work and AI training for ecommerce teams, so we have an interest in you taking this seriously. We have no affiliate, partner or paid relationship with Anthropic; we pay for Claude like anyone else, and we also use other assistants for tasks they do better. Product features are described as we use them; plans, limits and pricing change, so Anthropic's support center is the source for current specifics.
If you want the comparison first, the Claude vs ChatGPT vs Gemini comparison covers why Claude runs our production and where the others win. This post assumes the choice is made.
The Setup That Makes Workflows Repeatable
Everything below assumes four things exist. They take an afternoon to build and they are the difference between a chat that helps once and a system that runs weekly.
A Project holds standing instructions and files that every chat inside it sees. Ours contain: a rules file (banned words, claims policy, voice), a product facts file (ingredients, dimensions, variants, prices), the current listing copy, and the brand style guide. New chat, full context, no re-explaining.
The most valuable document in the Project is the list of things Claude must never do: never write a health claim, never invent a statistic, never use these twelve words, never change a price, always output American English, always flag anything it is unsure of. Rules that say no are followed more reliably than rules that say do.
Search term reports, business reports, settlement files, review exports and the inventory report, exported on the same day each week to a shared folder Claude can read through a connector or an upload. Workflows 4, 5, 11 and 14 are only repeatable because the input is the same file every week.
A skill is a folder of instructions and checks Claude loads for a specific job. We have one for blog production, one for scripts, one for listing audits. The skill carries the format, the quality checks and the pitfalls we have hit before, so the output does not drift between runs.
Workflow, as used here. A task with a fixed input (an export, a URL, a transcript), a fixed context (the Project's rules and facts), a fixed prompt, and a fixed check before the output is used. Something you can hand to a new team member with one sentence of instruction. If any of those four is improvised each time, it is a chat, and chats do not compound.
The MCP for ecommerce brands guide covers connecting Claude to the systems that hold your data, which turns several workflows below from "upload the export" into "read the report."
1-2: Review Mining And The Listing Rewrite
Workflow 1: Review mining across competitor ASINs
Input: the exported reviews of the top five competing ASINs (one to three stars and five stars, separately). Output: a table of objections and delights, ranked by frequency, each with three verbatim quotes.
The last sentence is the point. The mined objections become the "not X" headlines in A+ and the specific answers in the bullets. The review mining guide covers the method in full; this prompt is the version that runs in twenty minutes.
Workflow 2: The listing rewrite from mined objections
Input: the two tables from workflow 1, the current listing, the product facts file. Output: title, five bullets and description, each bullet mapped to one objection or delight.
The audit paragraph at the end catches the invented claim before it goes live. In a year of running this, that paragraph has flagged something about one time in five. The high-converting listing guide is the human standard the output is checked against.
3-4: A+ Module Mapping And Search Term Triage
Workflow 3: A+ module mapping
Input: the objection tables and the module list for the category (which modules, in which order, from the A+ teardowns on this site). Output: a module plan, one objection per module, with headline pattern and the fact each module carries, plus an A+ compliance pass.
Workflow 4: Search term report triage
Input: last week's Sponsored Products search term report. Output: three lists: terms to negate, terms to promote to exact match, and terms with spend and no sales that need a decision. This is the workflow that made the weekly PPC hour into a fifteen-minute review.
"Do not recommend bid changes" is deliberate: the triage is a data sort, and bidding is a judgment we keep. The PPC strategy guide covers what happens to the three lists.
5-6: The Weekly Business Review And Email Flows
Workflow 5: The weekly business review
Input: the Business Report by ASIN, the advertising summary, the inventory report and the settlement summary, all for the same seven days, plus the same four from the prior week. Output: a one-page narrative of what changed and why, with every number traceable to a file.
The "cause not visible" instruction stops the narrative from inventing explanations, which is the failure mode of every AI business summary. When it says cause not visible, that is the thing to go look at.
Workflow 6: Email flows in brand voice
Input: the flow brief (welcome, abandoned cart, post-purchase), the brand voice file, three past emails that performed well. Output: subject lines, preview text and body for each email, with the merge fields and the send timing.
Email is where brand voice drifts fastest, which is why the three past emails are attached every time rather than described. The email list building guide covers the flows themselves.
7-8: SOPs From Recordings And Support Macros
Workflow 7: SOPs from a screen-recording transcript
The single highest-return workflow on this list for a small team. Record yourself doing the task once while narrating (a shipment creation, a Vine enrollment, a coupon setup), transcribe it, and hand Claude the transcript.
The "usually" flag is the trick. Every operator has ten steps they do by feel; the transcript captures the feel and Claude turns it into a rule or a question. Ten recordings produced our onboarding manual.
Workflow 8: Customer service macro drafting
Input: the last 100 customer messages exported from the help desk, plus the return policy and the facts file. Output: the fifteen most common situations, a macro for each, and a list of situations that should never get a macro.
Macros save the team time. The "never a macro" list saves the brand. Both come out of the same export.
9-10: Browser Teardowns And Supplier Drafts
Workflow 9: Competitor listing teardown in the browser
With Claude's browser extension open on a competitor's live listing, this is a structured read of the page rather than a screenshot guess. Input: the open listing. Output: a teardown against our own listing's structure.
"Do not guess" matters here because a browser read is only as good as what rendered. If the A+ did not load, the answer should say so. The competitor spy tools roundup covers the paid tools that pull the numbers a page read cannot.
Workflow 10: Supplier negotiation drafts
Input: the supplier's last quote, our target, the volume history, the two things we want (price, terms, MOQ, lead time). Output: a first draft that asks for one thing and offers one thing.
The last instruction is the one that improved response rates. Most supplier email goes to people reading in a second language, and Claude is good at plain English when told to be.
11-12: Reorder Math And Creative Briefs
Workflow 11: Reorder point check
Not the calculation itself, which lives in a sheet. The check. Input: the inventory report, the last 90 days of unit sales by ASIN, lead times and safety stock per SKU from the facts file. Output: a list of SKUs that will hit the reorder point in the next 14 days, with the math shown.
"Do not estimate" is why this is a check and not a calculator. A missing lead time becomes a question rather than a guess. The reorder point guide has the formula the check runs.
Workflow 12: Creative brief and shot list
Input: the objection tables, the listing plan, the product. Output: a shoot brief with a shot list ordered by the listing image slots, each shot tied to the objection it answers.
The brief that ties every shot to an objection is the one the photographer can execute without a call. The creative brief template is the structure the prompt fills.
13-14: Script Variants And Settlement Reconciliation
Workflow 13: Short-form script variants
Input: one script that worked, the hook library in the Project, the product facts. Output: five variants that change one thing each. We run this through a scripting skill that carries the voice rules, so the output sounds like us rather than like a template.
Workflow 14: Settlement reconciliation
Input: the settlement report and the orders report for the same period. Output: every fee line summed by type, the effective rate of each against revenue, and anything that does not tie out.
This is the workflow that finds the storage surcharge you did not know started, the reimbursement you were owed, and the "other" line nobody can explain. It replaced a monthly hour of spreadsheet work with ten minutes of reading. The Amazon P&L guide covers what to do with the percentages.
15: Content Production With Skills
The post you are reading was produced this way, so this is the most first-party section in it. Our blog production runs through a skill: a folder holding the house template, the editorial standards, the verified link list, the deployment scripts and a written record of every mistake a previous batch made. Claude loads it, researches the calendar entry, writes to the template, runs the verification script (links, schema, meta lengths, mobile width, banned terms), and hands back a file we review and edit. The edits go back into the skill.
The mistakes file is the part that makes this compound. The last batch drafted in British spelling; the file now says so, and this batch did not. A previous batch stamped one post's code label onto every post's code blocks through a shared stylesheet; the file explains why, and it has not recurred. A skill without a mistakes file is a template. A skill with one is a colleague who remembers.
The output of workflow 15 is what earns the AI assistant traffic we measured in the GA4 post earlier in this series, which is the loop: Claude writes the content that the assistants, including Claude, then cite. The compounding content guide covers the strategy; the topical authority guide covers how the posts are clustered.
The Ecom Profit Box
Our library of ecommerce growth guides, including the AI operations and content frameworks behind this post.
Get It FreeSet Up Your Brand's Project
We build the Project, the rules file and the first three workflows with your team on a call, then you own it. Bring your exports and your listing.
Book A CallHow We Prompt: The Pattern Under All 15
Read the fifteen prompts again and the same six moves appear in every one.
- Name the input and where it is. "The attached report," "this listing," "the facts file in the Project." Claude works better when it knows which file is the source of truth for each fact.
- Specify the output as a shape. Three tables, a one-page review with five sections, five variants labeled. A shape is checkable; "help me with" is not.
- Constrain with negatives. Do not estimate. Do not recommend bid changes. Do not paraphrase quotes. Do not guess what is not on the page. Negative constraints are the most reliably followed instructions we have found.
- Demand the audit trail. "Cite the file and column." "List every claim and its supporting line." "Show the arithmetic." The trail is what makes the output usable without re-doing it.
- Make uncertainty an output. "Say cause not visible in data." "Flag anything you could not support." "Skip the SKU and say which input is missing." An assistant that is allowed to say it does not know stops inventing.
- Point at the rules. "Apply the rules file." The rules live in the Project, so the prompt does not restate them, and they are the same every time.
None of this is clever. It is the same discipline you would use briefing a capable new hire who has no context and no fear of being wrong. The founder AI stack covers the other tools around Claude that the exports come from.
The Four Things We Stopped Trusting It With
Every workflow above has a human check before the output does anything. These four are where the check is non-negotiable, because each one has cost us or a client something.
| Never Alone | What Happened | The Rule Now |
|---|---|---|
| Any number that ends up in a price | A landed-cost model with a unit conversion error, caught before it reached a price sheet, would not have been caught after | Claude shows the arithmetic; a person recalculates one line by hand before anything is priced |
| Anything it says a marketplace policy says | A confident, wrong statement of an Amazon fee threshold, from training data a year stale | Policy claims are checked against the live help page, every time; the rules file says "never state a policy without a URL" |
| Anything about a live listing it did not just read | A teardown of a competitor's "current" A+ that was two redesigns old, from memory rather than the page | Listing facts come from a browser read or an export in the same session, or they are marked unverified |
| Anything that goes to a customer | A macro that was correct and cold, sent to someone who had written a paragraph about their kid's rash | Every customer-facing message is read by a person before it sends, and the "never a macro" list from workflow 8 is enforced |
The pattern: Claude is reliable on what it can read in the session and unreliable on what it remembers. Every failure above was a memory problem. Every fix was to put the truth in the session. The why AI agents fail guide makes the same point about autonomous agents, where the stakes are higher because nobody is reading.
Where To Start If You Have An Hour
The honest counterweight, since we sell the setup: the setup is not hard, and most operators can build it from this post in an afternoon without us. What we add is the mistakes file from a year of doing it and the skills that carry it. If you would rather learn by doing, do; the fifteen prompts are the same either way.
The prompt is the least important part. The export it reads, the rules it obeys, and the check it runs against are what make the same task come out the same way every week.
What To Remember
- Build one Project per brand with a rules file, product facts and standing exports before writing a single prompt; that is what makes a task repeatable.
- The rules file works best as a list of "never" statements: never invent a statistic, never write a health claim, never state a policy without a URL.
- The highest-return workflows for a small team are review mining, search term triage, the weekly business review from exports, SOPs from narrated screen recordings, and settlement reconciliation.
- Every prompt names its input, specifies the output shape, constrains with negatives, demands an audit trail, and makes "I cannot tell from this data" an allowed answer.
- Skills with a mistakes file compound; a skill without one is a template, and the last batch's errors are the next batch's rules.
- Claude is reliable on what it reads in the session and unreliable on what it remembers; every failure we have had was a memory problem fixed by putting the truth in the chat.
- Four things never ship without a human: numbers that become prices, marketplace policy claims, statements about listings it did not just read, and anything sent to a customer.
Where This Came From
- Evolve Media Agency and Bear Basics Co. daily operations, September 2025 through September 2026. First-party; the fifteen workflows, prompts, setup and failure table are ours.
- Anthropic, Claude product overview and Claude support center, for current descriptions of Projects, file uploads, connectors and the browser extension; and Claude Code documentation for the command-line tool that runs our deployment scripts. Features, plans and limits change; these pages are current where this post is not.
- Inference note: the "one in five" flag rate in workflow 2 and the time savings quoted in workflows 4 and 14 are our own rough tallies from a year of use rather than measured benchmarks. No affiliate or partner relationship with Anthropic exists.