Claude Projects keep briefs, brand rules, and prior notes in persistent context. Pairing a Project with the Trends MCP connector turns that workspace into a standing trend desk: the same keyword list, growth windows, and live feeds on every session without re-uploading spreadsheets.
Claude Projects bundle persistent instructions, files, and chat history around one initiative. Trends MCP supplies live measurements from Google, TikTok, Reddit, Amazon, news, npm, Steam, and ranked platform feeds. Used together, a Project becomes a standing trend desk: the analyst reopens the same workspace each Monday, runs a fixed prompt list, and gets fresh numbers without re-explaining the category map.
Connector installation is documented on Trends MCP for Claude and the narrower Claude.ai OAuth connector guide. This page covers operating rhythm inside a Project.
A standard Claude.ai chat with Trends MCP enabled can answer a single question well. The context resets when the thread ends. Projects solve three recurrence problems for trend work.
Stable vocabulary. Sector labels, competitor spellings, and banned phrases live in Project instructions. Claude applies them on every message, which keeps briefs aligned with brand language.
Watchlists without re-uploading. A CSV or markdown file attached to the Project holds keywords, ASINs, subreddit names, or npm packages. The analyst edits one file when the list changes.
Repeatable prompts. Weekly pulses use the same sentence structure. The model already knows to call TrendsMCP, which sources to check, and how to format output.
Without a Project, teams re-paste watchlists or re-write rules each session. That friction pushes people back to static screenshots from dashboards.
| Asset | Contents | Update frequency |
|---|---|---|
| Project instructions | Role, sectors, growth windows, output format, citation rule ("cite TrendsMCP pulls with date") | Monthly or when strategy shifts |
watchlist.md |
Keywords, competitors, hashtags, subreddits | Weekly |
sectors.md |
Mapping from keywords to business units | Quarterly |
brief-template.md |
Headings for the Monday note | Rarely |
| Saved outputs | Prior weekly briefs the team wants to compare | After each run |
Keep instructions under roughly 2,000 words. Long instruction blocks compete with file context limits. Put bulky tables in watchlist.md.
You are the trend desk for [Company] consumer health products.
Always call TrendsMCP for numbers. Do not guess search volume.
Default growth windows: 3M, 12M, YTD on google search.
Add tiktok and amazon when the keyword is a consumer product.
Read keywords from watchlist.md. Process every row each Monday.
Output sections: Executive pulse, Movers up, Movers down, Live feeds, Open questions.
When data is missing, state the error code instead of improvising.
watchlist.md.Chats started outside the Project do not inherit Project files. Chats started before the connector was enabled may need a fresh thread.
Prompt:
Tool calls (one per feed):
get_top_trends(type='Google Trends', limit=10)
get_top_trends(type='TikTok Trending Hashtags', limit=10)
get_top_trends(type='X (Twitter) Trending', limit=10)
Three requests total. The overlap table becomes the opening paragraph of the weekly brief.
Prompt:
If the watchlist has twelve keywords and two sources, budget up to twenty-four requests unless batched. Some teams split across two days to stay inside the free tier.
Pick the top three movers from Step 2.
Prompt:
Two requests per keyword when using two sources.
Prompt:
Save the markdown output back into the Project if week-over-week comparison matters.
A simple markdown table keeps parsing reliable:
| keyword | sector | sources | notes |
|---|---|---|---|
| electrolyte powder | hydration | google search, amazon, tiktok | flagship category |
| creatine gummies | fitness | google search, amazon | new SKU line |
| wallstreetbets | finance signal | reddit | sentiment only |
Reddit rows use subreddit names without r/. npm rows belong in a separate table because package names are case-sensitive.
| Monitor shape | Approximate monthly requests |
|---|---|
| 10 keywords × Google only × weekly growth | 40 |
| 10 keywords × 3 sources × weekly growth | 120 |
| 3 live feeds × weekly pulse | 12 |
| 3 deep dives × 2 sources × weekly | 24 |
A full multi-source monitor can exceed the free 100-request tier in one month. Starters at 1,000 requests per month ($19) cover most single-team desks if deep dives stay focused.
Teams should document expected burn in the Project instructions so new members do not add redundant pulls.
Projects excel at interpretation, MCP at measurement.
Upload earnings call excerpts, campaign calendars, or retailer reset dates into the Project. Ask Claude to compare those dates against TrendsMCP series:
The calendar file is static context. The series is live. That split keeps narratives grounded without storing stale numbers in instructions.
Marketing wants bullet movers and creative angles. Ask for hooks tied to TikTok hashtags that appeared in get_top_trends.
Product wants npm or Steam signals. Add a section that pulls get_growth for package names or game titles from the watchlist.
Investor relations wants YTD language. Standardize on percent_growth=['YTD','12M'] in instructions so every brief uses the same windows.
Agency clients may need JSON attachments:
Claude can paste JSON into a code block for export even though Projects do not auto-sync to S3.
No silent history. MCP does not write into Project files automatically. Someone must save outputs if history is required.
Connector off by default in old chats. New weekly threads avoid missing tools.
Duplicate pulls cost money. Re-running the same keyword and source within a session still counts per Trends MCP billing rules.
404 on niche terms. Thin Amazon or Steam series return not_found. Instructions should tell Claude to report the error and move on.
Plan gates. Browser Projects need Claude.ai Pro, Max, or Team for remote MCP. Desktop Projects need local MCP config instead.
FAQ