Make the handoffs explicit before automating the pipeline
A small export studio is considering garden-tool organizer bags for overseas distributors. One person handles product selection, listings, inquiry replies and customer development. This is a fictional working example: no stock has been bought, listings published or messages sent. The objective is to separate established facts, assumptions and the input required by the next stage, so fluent AI copy cannot disguise missing evidence.
The source covers three substantial workloads: product research, multilingual listings, and outreach with diligence. This guide preserves those stages without repeating unverifiable download counts or promises to learn most of a customer's background in seconds. An assistant can organize accessible material. It cannot manufacture market data when no data source exists, and locating a website does not complete a commercial credit assessment.
Research three candidates on comparable terms
Compare organizer bags, kneeling pads and plant labels. Specify destination market, sales channel, price band and observation period first. Every observation needs a source URL, access date, currency, tax and shipping basis, and a description of what was observed: an asking price, sales figure, review count or another signal. Marketplace rank and reviews may inform a hypothesis; neither independently measures the size of an entire market. Check product variants so a multi-pack price is not compared with a single item.
Build a candidate matrix covering demand evidence, competing samples, acquisition cost, packed dimensions, shipping constraints, minimum order quantity, replenishment time and evidence gaps. Only measured samples or confirmed supplier specifications belong in the facts column. Show the assumptions behind a margin scenario, including returns and exchange-rate sensitivity, rather than presenting a precise-looking guaranteed-profit score. Retain conflicting sources and assign a follow-up instead of letting the model choose the more confident wording.
When an authorized local CSV export already exists, use Exploratory Data Analysis at this input-quality step. Its supported workflow can inspect missingness, anomalies and aggregate distributions. Check duplicate SKU observations, prices containing mixed currency strings and the number of usable observations for each candidate before interpreting a comparison. It does not collect marketplace data, estimate an unobserved market size or establish a causal sales forecast.
Confirm the permitted data root and file format before running anything. Core CLIs require Python 3.11 or later; optional formats have additional dependencies documented at the pinned source. Preserve the original export and record cleaning rules separately. Do not silently discard unusually high prices that may actually represent bundles. Deliver a quality report, comparable sample table and missing-information list. Those support a human sampling decision; a high generated score must not automatically place an order.
Draft listings from a versioned product fact card
Create a fact card for the organizer bag: measured dimensions, material, colors, package contents, supported tool sizes, care instructions and evidence locations. An unknown waterproof rating remains unknown. Wipe-clean fabric does not justify a waterproof claim. If the photograph shows a bag and several demonstration tools, make clear which items are actually included. Certification, environmental and safety claims need applicable evidence for this specific product rather than plausible marketing language.
Supply target language, audience, platform, facts, prohibited claims and unresolved questions. Request a title, relevant keyword candidates, selling points and a description as separate outputs. Lead the title with the product category and a real distinction. Select keywords for fit rather than repetition. Review each language for units, idiom, tone and search intent; a literal translation can preserve every word while missing the local customer's meaning.
A concrete drafting request is: “Use fact card v3 for an English distributor catalogue page. Preserve every measurement and package quantity. Put unknowns in editor notes. Add no certification or delivery-date promises. Return a title, four selling points and a short care paragraph.” Ask for a fact-to-copy mapping alongside the draft: which statements derive from which evidence, which are merely wording choices, and which require confirmation.
A reviewer then checks the current channel requirements for length, restricted claims, imagery, attributes and any relevant AI-content handling. This article does not declare a universal character limit or certify compliance across marketplaces. Inspect a preview for truncation, wrapping and variant selection. Save the product fact version, language revision, reviewer and publication state. When a specification changes, revisit both languages. The deliverable at this stage is an approved draft and review record, not a claim that generating text has put a product on sale.
Answer inquiries without inventing commercial commitments
A prospect asks whether customization is available, what minimum quantity applies and whether delivery can happen next week. Extract answers from approved commercial information and route unknown lead times to the supplier. Structure a reply around acknowledgement, answers to the actual questions, explicit pending items and one next step. A casual “of course” should not accidentally promise a price, brand-use permission and a shipping deadline together.
Separate sample, production and repeat-order conditions. A quotation needs currency, validity, quantity basis and a clear description of included expenses. If the trade term or destination requirement has not been agreed, identify it as a variable rather than allowing a generator to fill in contractual defaults. Compare the language versions for quantities, dates and attachment names. Check the recipient and real attachments before sending. No inquiry response or quotation was transmitted as part of writing this tutorial.
Develop customers through sourced evidence cards
Start from product use and target distribution channel. For each business lead, record the official site, operating region, evidence of category fit and a public business contact route. Similar company names are common. Match combinations such as domain, address and registration identifier before attaching another organization's news or history to the prospect. Prefer permitted sources and public commercial channels over accumulating the largest possible contact list.
Split a diligence card into the company's own statements, independent evidence and unresolved questions. A website's claim of many years in business is self-description, not verified registration age. An unavailable page or unanswered call does not establish fraud. Registration, payment capacity and any transaction-specific screening need appropriate sources and review for the actual market and deal. An AI summary is neither a formal credit report nor a guarantee of safe trading.
For the garden distributor, a useful match might be its published catalogue of hand tools and storage accessories. The outreach draft can mention that verifiable connection. It must not fabricate knowledge of a private purchasing plan. Offer one concrete value, a brief accurate product fact and a low-burden question. The owner reviews the draft, applicable marketing requirements, opt-out handling and sending authority. This workflow creates a draft; it does not authorize private-contact harvesting or bulk sending.
Track discovery, verification, draft approval, sending and response as separate states. If an address changes, retain the evidence date and verify it before use. Do not count a drafted email as a completed contact or an unverified business as a qualified buyer. A small, relevant and sourced list is more actionable than a large spreadsheet whose identities cannot be reconciled.
Attach a Skill to the precise work it can perform
There is no requirement to install two tools at every stage. The verified SkillPI EDA listing fits local input-quality inspection during research. Its detail page provides the pinned source, MIT license evidence and runtime prerequisites. It does not perform customer credit investigation, certify listing compliance or send outreach. The other package names in the source have not been established here as verified, purpose-matching SkillPI records, so this guide does not invent links for them.
For a new candidate, inspect its source and SKILL.md, data destinations, account requirements, costs and expected artifacts. Try one non-sensitive sample before allowing a wider workload. Discovery, installation, invocation and accepted output are different states. Uninstalling a package cannot recall sent messages, remove every external authorization or reverse files it has already written. Retest the same sample after an update and revisit older article associations when a package's scope changes.
GitHub research found FeedGen, a project concerned with generative Shopping-feed optimization and review. It is relevant as an inspectable example of structured product inputs and approval before output, rather than a universal export-sales solution. Its account and service dependencies may be unnecessary for a small pilot. This example therefore uses fact cards and explicit human gates without installing that integration or connecting a commercial account. Adoption would require checking the selected revision, license, dependencies and the current channel rules separately.
Troubleshooting and frequently asked questions
What if the three products lack comparable data? Narrow the market and observation window, report missing fields and collect evidence; do not fill gaps with invented totals. What if the draft resembles a competitor? Return to the studio's own facts and rewrite, excluding distinctive borrowed language and unlicensed images. What if translated specifications disagree? Compare by SKU and fact-card version, freeze the faulty draft and preview both revisions after correction.
What if diligence sources conflict? First rule out mistaken company identity, then preserve source dates and uncertainty. What if outreach receives no replies? Delivery, recipient relevance, timing and the offer can all matter. Record delivery and replies separately rather than concluding that a few messages prove product demand is absent. Can one person operate the whole workflow? Staged assistance can reduce repetitive organization, while purchasing, quotations, rule interpretation and business relationships still need an accountable owner.
The reviewable handoff is a candidate matrix, input-quality report, bilingual listing draft, inquiry template and sourced customer cards. Label every unexecuted stage as pending. This creates a practical path from research to a considered contact without confusing generated text with measured demand or completed business actions.
Sources
The English source supplies the three business stages; this guide uses an independent garden-supply example. FeedGen is the GitHub reference for a structured feed-generation and review approach, with its own integration requirements. The pinned EDA source supports the bounded local exploration step. None of these references establishes a market-size estimate, campaign result or completed transaction for this example.