TL;DR
A Claude skill is a SKILL.md file that teaches Claude one repeatable job. For a product feed, six skills cover the recurring work: Merchant Center audit, Shopping titles, product_type and custom labels, Meta catalog quality, catalog ad fatigue and competitor catalog ads. They run on an export; past 10,000 products you need rules, tests and rollback.
You export your Merchant Center feed on Monday, paste 2,000 rows into Claude, and ask why 340 products are disapproved. The answer is good. On Thursday you ask the same question with a fresh export and get a different structure, different thresholds and a missing column. A Claude skill fixes that: you write the method once, and Claude applies it the same way to every export.
Most published Claude skill libraries for marketers cover ad account performance, Shopify or CRO. Few treat the product feed as the subject, even though the feed decides what Google Shopping, Performance Max and Meta Advantage+ catalog ads can show. This guide gives six feed and catalog skills you can copy, the input each one needs, the MCP servers that let Claude read a live feed, and the point where a skill on a CSV stops being enough.
| # | Skill | Problem it handles | Input | Cadence |
|---|---|---|---|---|
| 1 | merchant-center-feed-audit | Disapprovals, missing attributes, invalid GTINs | Merchant Center product export + diagnostics | Weekly |
| 2 | shopping-title-rewrite | Titles that hide brand, type and key attributes | Product export with category and attributes | Per category, then on new products |
| 3 | product-type-custom-labels | Campaigns you cannot segment by margin, season or stock | Product export + margin or sales data | Monthly |
| 4 | meta-catalog-quality | Catalog items with issues, product sets too small to deliver | Meta data feed file + Commerce Manager issue report | Weekly |
| 5 | catalog-ad-fatigue | Catalog ad CTR sliding while frequency climbs | Meta ad-level export, daily breakdown | Weekly |
| 6 | competitor-catalog-ads | No view of competitors’ catalog ad formats and offers | Notes from the Meta Ad Library | Quarterly |
Skill, prompt or MCP: what is the difference for a product feed?
A Claude skill is a folder containing a SKILL.md file: YAML frontmatter with a name and a description, then the instructions. A prompt is a one-off request. An MCP server is a connector that gives Claude access to a system, such as Merchant Center or a Meta ad account. For feed work, the skill holds the method, the prompt triggers it, and the MCP decides whether Claude reads a CSV or the live feed.
| Prompt | Skill | MCP server | |
|---|---|---|---|
| What it is | A request in the chat | A SKILL.md file with instructions, optional scripts and reference files | A connector to an external system |
| What it brings | One answer | The same method on every run | Live data and, depending on the server, actions |
| Feed example | ”Why are these products disapproved?” | A fixed audit against Google’s product data spec | Reading product statuses from Merchant Center |
| Limit | Output drifts from one run to the next | Works on what it receives | Reads or writes only what its permissions allow |
According to Anthropic’s Agent Skills documentation (consulted November 2026), Claude loads skills in stages. The name and description of every installed skill stay in context, at about 100 tokens each. Claude reads the body of the SKILL.md only when your request matches the description, and opens extra files or runs scripts only when the instructions call for them. You can install dozens of skills without filling the context window.
Two consequences for a feed team. The description decides when the skill fires, so it should use the words your team types (“disapproved”, “GTIN”, “custom label”). And a skill can ship a script: a Python file that checks GTIN check digits on 50,000 rows is more reliable than asking the model to do it in its head.
How do you install a Claude skill?
You install a skill by placing its folder where Claude looks for it. In Claude Code, that is ~/.claude/skills/ for you alone or .claude/skills/ inside a repository for the whole team. On claude.ai, you upload the folder as a zip from the skills settings, on paid plans with code execution turned on. Through the API, skills run with the code execution tool.
~/.claude/skills/merchant-center-feed-audit/SKILL.md
~/.claude/skills/merchant-center-feed-audit/scripts/check_gtin.py
.claude/skills/shopping-title-rewrite/SKILL.md
.claude/skills/shopping-title-rewrite/reference/title-rules.md
Three rules from Anthropic’s documentation matter here:
nametakes 64 characters at most, in lowercase letters, numbers and hyphens, and cannot contain “anthropic” or “claude”.descriptiontakes up to 1,024 characters and must say what the skill does and when to use it.- Skills do not sync between claude.ai, Claude Code and the API. You install them on each one.
Anthropic recommends using only skills you wrote yourself or got from Anthropic, and auditing any other skill (SKILL.md, scripts, resources) before use, since a skill can call tools and run code. Read every file of a skill you download from GitHub before you point it at an account that holds your catalog and your ad spend.
Skill 1: how can Claude audit a Merchant Center feed?
Claude can audit a Merchant Center feed when you give it the product export and the diagnostics, and when the skill checks each product against Google’s product data specification rather than general knowledge. The skill flags missing required attributes, invalid identifiers and values over the limits, groups products by issue, and ranks issues by the number of products affected. It explains the fix; your feed tool or team applies it.
Input. The product export from Merchant Center (or your source feed) with at least id, title, description, link, image_link, availability, price, brand, gtin, mpn, google_product_category, product_type, plus the diagnostics export listing item issues.
Checks taken from Google’s product data specification (consulted November 2026):
titleis required, 150 characters maximum;descriptionis required, 5,000 characters maximum.gtinaccepts valid GS1 values only (UPC, EAN, JAN, ISBN, ITF-14) with a correct check digit. Google states that an incorrect GTIN leads to disapproval, and asks you never to guess one.brandis required for new products except movies, books and music recordings, 70 characters maximum.mpnis required only when the product has no GTIN.identifier_existsset to “no” is only for products without unique identifiers.priceneeds a number and an ISO 4217 currency code, and must match the landing page.image_linkmust be crawlable. Google lists a 500 x 500 pixel minimum enforced from January 31, 2027, which is worth flagging now.
---
name: merchant-center-feed-audit
description: Audits a Google Merchant Center product export against Google's
product data specification. Use when asked why products are disapproved,
which attributes are missing, whether GTINs are valid, or to prioritize
feed fixes. Not for Meta catalogs.
---
## Steps
1. Echo the row count, the columns found and the columns missing. Stop and
ask for a re-export if id, title, price, availability or link is missing.
2. Run scripts/check_gtin.py on the gtin column. Report invalid check digits.
3. Check each attribute against reference/spec.md (limits and rules).
4. Join the diagnostics export on id. Group products by issue.
5. Rank issues by products affected, then by share of clicks if a
performance column is present.
## Output
Table: issue | attribute | products affected | example ids | fix | source rule
Then the 3 fixes to make first and the products to exclude rather than fix.
Never invent a GTIN, a brand or a price.
Keep the spec rules in reference/spec.md with the date you copied them. When Google changes a rule, you update one file. For the full list of attributes and common rejection causes, see our Google Merchant Center guide and the list of Shopping feed errors and their fixes.
Skill 2: how do you rewrite Google Shopping titles with Claude?
You rewrite Shopping titles with Claude by fixing the pattern per category first, then applying it product by product from attributes already in the feed. The skill puts the terms shoppers search at the start, keeps the brand and the distinguishing attributes, stays under 150 characters, and refuses to add a fact that the product data does not contain. You test the new titles on a share of the catalog before rolling them out.
Google’s specification asks for titles that match the landing page, avoid promotional text and text in all capital letters, and include the attributes that distinguish variants, such as color or size. A pattern per category turns that into something Claude can apply the same way on every row:
| Category | Pattern | Example |
|---|---|---|
| Apparel | Brand + product type + gender + material + color + size | [Brand] men’s organic cotton sweatshirt, navy, M |
| Auto parts | Brand + part + vehicle compatibility + reference | [Brand] wiper blades for [model], 600/400 mm |
| Furniture | Product type + material + dimensions + color + brand | Oak coffee table 120 x 60 cm, natural, [Brand] |
The examples show the structure only. Every value must come from your feed.
---
name: shopping-title-rewrite
description: Rewrites Google Shopping product titles from feed attributes using
a pattern per category. Use when asked to optimize, rewrite or standardize
Shopping titles. Never adds a fact absent from the product data.
---
## Steps
1. Load the pattern for the product's category from reference/title-rules.md.
If no pattern exists, propose one and stop for approval.
2. Build the title only from columns present in the row.
3. Reject the title if it exceeds 150 characters, repeats a word, contains
promotional text ("free shipping", "sale") or full capitals.
4. Keep the original title in a column next to the new one.
## Output
CSV: id | original_title | new_title | pattern | attributes_used | warnings
Then the count of products skipped for missing attributes, by attribute.
The before and after columns matter more than the rewrite itself. They let you split the catalog and measure. Feu Vert Spain enriched its feed with AI-generated attributes and measured +32% ROAS in a 50/50 catalog A/B test, with -15% CPC on Google Shopping. The attributes guide lists the fields worth filling before you touch titles. Past a few categories, Feed Enrich generates titles and attributes product by product from the same patterns, with batch generation for large catalogs.
Skill 3: how should you set product_type and custom labels?
You set product_type as your own full category path and use the five custom labels to segment campaigns by what the feed does not say on its own: margin, season, stock depth, best-sellers. Google’s specification allows 750 characters for product_type, uses only the first value for bidding and reporting, and limits each custom label to 100 characters and 1,000 unique values across the account.
That 1,000-value ceiling is where hand-made labels break. A label filled with product names or exact prices runs out fast. The skill proposes buckets instead:
| Label | Use | Example values |
|---|---|---|
| custom_label_0 | Margin band | margin_high, margin_mid, margin_low |
| custom_label_1 | Season | ss26, aw26, permanent |
| custom_label_2 | Performance | best_seller, zombie, new |
| custom_label_3 | Stock depth | stock_deep, stock_low |
| custom_label_4 | Price band | price_0_25, price_25_75, price_75_plus |
---
name: product-type-custom-labels
description: Proposes product_type paths and custom_label_0-4 values for a
Merchant Center feed from catalog, margin and performance data. Use when
asked to segment Shopping or Performance Max campaigns, label products, or
fix product_type. Respects the 1,000 unique values per label limit.
---
## Steps
1. Read the category tree. Propose a product_type path per category,
full path from top level, with ">" separators.
2. Ask which business dimensions campaigns should split on. Default to
margin, season, performance, stock and price band.
3. Define thresholds per bucket from the data (quartiles), state them.
4. Count unique values per label. Refuse any label above 1,000.
## Output
Mapping table: id | product_type | custom_label_0 ... custom_label_4
Then the bucket definitions and the campaign split they enable.
Labels go stale. A best-seller in March can be a zombie in June. Run this skill monthly, or turn the buckets into a rule in Feed Enrich, the Dataïads product feed optimization tool, which recalculates them on each feed update without a manual export.
Skill 4: what should a Meta catalog quality check cover?
A Meta catalog quality check covers the products Meta flags with issues, the errors in each data feed upload, and product sets that end up empty or too small to deliver. The skill reads the data feed file you send to Meta, the issue report from Commerce Manager and your product set filters, and ranks problems by the number of products they remove from active product sets.
Input, from Meta’s Business Help Centre (consulted November 2026). In Commerce Manager, open your catalog, then Catalog > Data sources, select the feed and click Download issue report. Meta notes that the report shows a sample of the issues, and that the button is unavailable when the upload session is too old or the file contains more than 100,000 errors. For the full picture, add the Products tab filtered by issue (Filters > Issues > Advantage+ catalog ads) and your source feed file.
Meta’s help pages list the issues worth checking first: missing title, description, price, availability, product link, condition or image link, broken links, descriptions in uppercase or identical to the title, ID conflicts with the group ID, missing universal identifiers or invalid GTINs, duplicated IDs, and invalid prices or currencies.
Product sets have their own minimum. According to Meta’s product set troubleshooting page, a set needs at least four unique, valid products for the collection format and two for the carousel format, or the ads may pause. Variants of the same product count once, and products sharing the same image count as one. A set of ten products where eight are variants of one item holds three unique products. Filters drift too: after a catalog update, products that no longer match a set’s filters drop out of it.
---
name: meta-catalog-quality
description: Checks a Meta catalog from its data feed file, the Commerce
Manager issue report and product set filters. Use when asked why catalog
items are rejected, why a product set is small or empty, or whether the
Meta catalog is ready for Advantage+ catalog ads. Not for Merchant Center.
---
## Steps
1. Echo item count, columns and the date of each file. Say that Meta's
issue report is a sample, not the full list.
2. Check required fields: id, title, description, availability, condition,
price, link, image_link. Flag missing values per field.
3. Flag duplicate ids, ids equal to item_group_id, descriptions in
uppercase or identical to the title, invalid GTINs and price formats.
4. Group the issue report by issue type, with counts and example ids.
5. For each product set filter provided, count unique products: one per
item_group_id, one per distinct image, excluding items with issues,
archived or out of stock.
## Output
Table: issue | items | product sets affected | example ids | fix
Then product sets under the minimum (4 for collection, 2 for carousel)
or close to it, and the filter or issue that causes it.
We cover how the catalog interacts with campaign setup in our guide to Meta Advantage+ shopping campaigns. Feed Enrich publishes the same enriched feed to Meta Catalog and Merchant Center, so a fix made once reaches both.
Skill 5: can Claude detect creative fatigue in catalog ads?
Yes, from an ad-level export with a daily breakdown. A common sign of catalog ad fatigue is CTR sliding while frequency and CPM rise. In our experience it differs from single-image fatigue: the products rotate, but the template, overlay and framing stay the same for weeks. The skill flags ads and templates whose CTR dropped from their seven-day peak, and recommends a template or overlay change rather than new product shots. For the full method without Claude, read our guide to creative fatigue on Meta catalog ads.
Input. Meta Ads Manager, ads tab, last 30 days, breakdown by day: ad name, ad set, template or creative name, spend, impressions, reach, frequency, CTR, CPM, results, cost per result.
---
name: catalog-ad-fatigue
description: Flags fatigued Meta catalog ads and templates from an ad-level
export with a daily breakdown. Use when catalog ad CTR is falling, frequency
is rising, or when asked which catalog templates to rotate.
---
## Steps
1. Stop if the daily breakdown or frequency column is missing.
2. Per ad and per template: 7-day peak CTR, current 7-day CTR, drop in %.
3. Urgent: 3+ consecutive days of decline or drop above 30%.
Warning: drop of 10 to 30%. Otherwise healthy.
4. Group by template. A template fatigued across several ad sets needs a new
design; a single fatigued ad set points to the audience.
## Output
Table: template | ads | spend | peak CTR | current CTR | drop % | frequency |
status | action (new template, new overlay, audience change)
The thresholds are a starting point. Adjust them to your account history. Rotating templates by hand works for a few ad sets; at catalog scale you need templates generated per product, and a test to pick the winner. Before its sales period, Faguo ran an A/B test on its Meta catalog ads: the Smart Creative template showing price and promotion took 66% of the budget, with +86% impressions, +7% CTR and -5% CPC against the classic catalog. Creative variety also weighs more since Meta rolled out Andromeda, its ad retrieval engine, in December 2024; our article on visual signals in catalog ads covers the consequences.
Skill 6: how do you analyze competitors’ catalog ads in the Meta Ad Library?
You analyze competitors’ catalog ads by collecting their active ads from the Meta Ad Library, noting which ones look like product carousels or catalog formats, then asking Claude to compare formats, overlays, offers and freshness with your own. The skill works on what you collect. It does not scrape the Ad Library, and it marks every conclusion that rests on fewer than ten ads.
What the Ad Library shows. According to Meta’s Help Centre (consulted November 2026), anyone can search the ads currently active across Meta products: go to facebook.com/ads/library, select a country, choose All ads, search for the advertiser, and open See ad details. Results are grouped by Facebook Page. For commercial ads, you see active ads only; Meta keeps inactive ads for ads about social issues, elections or politics (seven years) and, per Meta’s transparency site, for ads delivered in the EU (one year after their last impression). Meta does not label an ad as a catalog ad, so you classify the format from what the preview shows.
Input. For three to five competitors: number of active ads, format of each ad as you read it (single image, video, carousel, product carousel), visible overlay (price, discount, badge), headline, CTA and the dates shown in the ad details. A table pasted into the chat is enough.
---
name: competitor-catalog-ads
description: Compares competitors' catalog and carousel ads collected from the
Meta Ad Library with our own catalog ads. Use when planning catalog creative,
or when asked what competitors show in their product ads.
---
## Output
Per competitor: active ads, share of catalog-style ads, overlays used,
offers shown, median ad age, new ads in the last 30 days.
Then: overlays and offers nobody uses, conventions everyone follows,
and 5 template tests for our next sprint, each tied to an observation.
Flag any conclusion based on fewer than 10 ads.
How do you move from a CSV to the live feed with MCP?
You move from a CSV to the live feed by connecting Claude to an MCP server that reads Merchant Center, the Meta catalog or your feed tool. The skills stay the same; the input changes from an attached export to live data. Read access is enough for skills 1 to 6. Write access changes what a mistake costs, so it needs approval steps and a change log.
| MCP server | What it reaches | Access documented | Source |
|---|---|---|---|
| Google Merchant API MCP service | Merchant Center data and insights | Alpha since May 2026, focused on read-only access | Merchant API release notes |
Meta Ads MCP server (mcp.facebook.com/ads) | Reporting, ad management, catalog creation and management, signals | Permission-scoped tools; catalog tools create catalogs, add product data and troubleshoot feed and item visibility | Meta developer documentation |
| Dataïads MCP server | Your feeds, Merchant Center, Google Ads and Meta data, Dataflow pipelines, Catalog Ads | Read and write within your Dataïads account; pipeline changes versioned in Dataflow, with rollback | Aïda page |
| Other feed tools’ servers and agents | The enriched feed, rules and exports | Depends on the vendor | Vendor documentation |
Three facts to keep in mind when you plan this:
- Google retired the Content API for Shopping on August 18, 2026 and started degrading the service on September 1, 2026. Scripts and connectors built on it need to move to the Merchant API.
- Google also publishes Merchant API Agent Skills for developers who build on the API. They help write integration code; they do not audit your feed.
- Meta opened its Ads AI connectors in open beta on April 29, 2026, and in July 2026 added MCP server rules that let anyone with full control of the business portfolio limit what agents can do, from budget changes to catalog updates. We looked at what this means for agencies in our article on Meta Ads AI connectors.
Before you give an agent write access to a catalog that feeds live campaigns, decide three things: who approves a change, where the change is logged, and how you roll it back.
How does Aïda run these skills on the full catalog?
Aïda is the AI agent of the Dataïads platform. You ask in plain language, Aïda reads your Merchant Center feed and your Google Ads and Meta performance, shows a plan before anything changes, and acts only after you approve it. It then builds the pipeline in Dataflow, publishes through the Merchant API, Meta Catalog or export files, and sends a before and after report on every run.
The six skills above map to requests you can type into Aïda as they are:
| Skill | Request to Aïda | What runs behind it |
|---|---|---|
| 1. Merchant Center audit | ”Run a full audit of my Merchant Center feed” | Full audit of the feed and its attributes, in Feed Enrich |
| 2. Shopping titles | ”Find under-exposed best-sellers and fix their feed” | Titles, descriptions and attributes generated product by product, with batch generation for large catalogs |
| 3. Custom labels | ”Where are my hidden gems and budget sinks?” | SQL, segmentation and clustering in Dataflow, replayed on every catalog change |
| 4. Meta catalog quality | ”Fill in missing colors and materials” | Enriched attributes published to Meta Catalog |
| 5. Catalog ad fatigue | ”Launch your first Catalog Ads across the whole feed” | Catalog Ads on your templates in Smart Creative, plugged into the enriched feed |
| 6. Competitor catalog ads | Stays a manual skill | Aïda works on your own data, not on competitors’ ads |
The difference with a skill on an export sits in what happens after the diagnosis. Aïda tests a preview on a sample before every save, keeps a history of every change with its author, and lets you roll back at any time. You can also ask Aïda to run a change on part of the catalog first, as an A/B test against the current feed, the way Feu Vert Spain measured its enriched feed on a 50/50 split.
Claude skill or Aïda: which one when?
Use a Claude skill to write and test a method on an export; use Aïda to run that method on the full catalog. A skill teaches Claude how to do a job. Aïda already has the data, the platform’s tools and the safeguards: it knows your catalog, keeps a memory per project, and acts in Dataflow after your approval. The two connect, and you do not have to choose one interface.
| Claude skill | Aïda | |
|---|---|---|
| What it is | A SKILL.md file: instructions, scripts, references | The AI agent of the Dataïads platform |
| Data | The export you attach, or what a generic MCP server exposes | Merchant Center, Google Ads, Meta Ads, Sheets, BigQuery and product feeds, joined in one project |
| Memory | None between sessions, outside its files | Standing instructions and a memory of past decisions per project |
| Domain knowledge | What you write into the skill | Dataïads docs and recipes, standard attributes, each channel’s requirements |
| Models | Claude | The models of the Dataïads platform for text, images and video |
| Output | A diagnosis, a recommendation, a file | A pipeline in Dataflow, published through the Merchant API, Meta Catalog or export files |
| Safety | What you build yourself | Plan before action, preview on a sample, approval, history, rollback |
| Scale | A sample or a category | The full catalog, replayed on every catalog change |
Aïda is not tied to one model. The Dataïads models page lists 60 models in production, among them Gemini and GPT-5 for text, Imagen, Nano Banana and Flux for images, Veo, Kling, Seedance and Sora 2 for video. Each task can run on a different model: rewriting 50,000 titles, generating lifestyle images for Catalog Ads, or turning a product photo into a video ad. A Claude skill stays within what Claude does on its own.
Three ways to connect your skills to Dataïads:
- Move the method into an Aïda project. Each Aïda project keeps standing instructions. The title patterns, label buckets and checks you wrote in a SKILL.md become the instructions Aïda follows on your live feed, and your method stays yours.
- Keep Claude or ChatGPT as your interface. Dataïads exposes its platform tools through an MCP server, so you can connect Claude or ChatGPT to your Dataïads account. Your skill keeps the method; the Dataïads tools read the full feed, build the pipeline and publish it, with changes versioned in Dataflow.
- Start with a skill, scale with Aïda. Run the skill on an export to find the problem and agree on the rule, then hand the rule to Aïda for the full catalog and every channel.
What breaks beyond 10,000 products?
A skill on a CSV breaks at scale for four reasons: the export no longer fits in one context window (a single description can run to 5,000 characters), the output of a run cannot be replayed on the next feed update, you cannot measure the change without a test design, and nothing undoes a bad batch. Our rule of thumb at Dataïads: below 10,000 products, a skill gives a good diagnosis. Past that, the skill should define the rules and a pipeline should apply them.
| Skill on an export | Feed pipeline (Dataïads) | |
|---|---|---|
| Volume | A sample or a category | The full catalog, every market |
| Repeatability | A new run on each export | Rules replayed on every feed update |
| Measurement | Before and after, by hand | A/B split on products, measured per metric |
| Safety | You review the file | Approval, versioning, rollback |
| Channels | One export per channel | One enriched feed published to Merchant Center, Meta and other channels |
Our customers’ catalogs show the gap. Castorama optimized a feed of 63,000 products and measured -4% CPC and +10% CTR on Google Shopping. La Redoute manages more than one million products, activates more than 200,000 SKUs in Meta catalog ads, and cut its creative tool costs by 30%. Neither job fits in a chat window, and both need the same rules applied again on the next update.
A practical split: use the skills to find the problem and write the rule, test the rule on a share of the catalog, then hand it to Feed Enrich, the Dataïads product feed optimization tool, which applies it to every product and publishes the same enriched feed to Merchant Center, Meta and your other channels.
Key takeaways
- Write the method once in a SKILL.md, with a description that uses your team’s words, so Claude applies the same checks to every export.
- Base feed skills on Google’s product data specification and keep the rules in a dated reference file.
- Start with the Merchant Center audit, then titles and custom labels, then the Meta catalog and catalog ad fatigue.
- Connect an MCP server for live data, starting with read access; add write access only with approval, logging and rollback.
- Past 10,000 products, let the skill define the rules and a feed pipeline apply them, with an agent such as Aïda handling the plan, the approval and the rollback.
FAQ
What are Claude skills?
Claude skills are folders containing a SKILL.md file with a name, a description and instructions, plus optional scripts and reference files. Claude keeps only the name and description in context and loads the rest when a request matches. Skills work in Claude Code, on claude.ai with code execution, and through the API, but you install them on each one.
Can Claude audit a Google Merchant Center feed?
Yes, from a product export and the diagnostics file, or live through an MCP server such as Google’s Merchant API MCP service, which is in alpha with a focus on read-only access. A skill makes the audit consistent by checking every product against Google’s product data specification. Claude explains the fixes; applying them happens in your source feed or feed tool.
Which MCP connects Claude to Merchant Center and the Meta catalog?
Google offers a Merchant API MCP service in alpha, focused on reading Merchant Center data and insights. Meta’s Ads MCP server, at mcp.facebook.com/ads, includes catalog tools that create catalogs, add product data and troubleshoot feed issues, within the permissions you grant. Feed tools also expose their own servers or agents.
How do I rewrite Google Shopping titles for 50,000 products with Claude?
Define one title pattern per category, write it in the skill’s reference file, and have Claude build titles only from attributes in the feed, under 150 characters. Test the new titles on part of the catalog first. At 50,000 products, run the pattern as a rule in a feed pipeline rather than in successive chat sessions.
Can Claude detect creative fatigue in Meta catalog ads?
Yes, from an ad-level export with a daily breakdown. A fatigue skill compares each ad’s current CTR with its seven-day peak, tracks frequency and CPM, and groups results by template, since catalog ads rotate products but keep the same template. A template fatigued across several ad sets needs a new design.
Can I connect Claude or ChatGPT to Dataïads?
Yes. Dataïads exposes its platform tools through an MCP server, so Claude or ChatGPT can read your feeds and your Merchant Center, Google Ads and Meta data, build pipelines in Dataflow and publish them, within your Dataïads account. You can also move the method of a Claude skill into the standing instructions of an Aïda project and work with Aïda directly.
Are Claude skills from GitHub safe to install?
Only after you audit them. Anthropic recommends using skills you wrote or got from Anthropic, and auditing any other skill before use, because a skill can run code and call tools. Check every SKILL.md and script. We also advise giving a downloaded skill read-only access to your ad accounts and feeds until you have tested it on an export.
When do I need a feed tool instead of a Claude skill?
In our experience, when the catalog passes 10,000 products, covers several markets or channels, or when changes must be replayed on every feed update, tested and reversible. A skill diagnoses and defines rules. A feed tool such as Feed Enrich applies those rules to the full catalog and publishes the result to each channel.
If you want to see what these checks find on your own catalog, start with a free audit of your product feed.
