Get a weekly Claude Code report that links shipped pull requests to KPI changes and scores the likely impact
A product lead connects Claude Code to Metabase and GitHub so each week it compares 5 to 10 KPIs with last week, maps the week's pull requests to them with a confidence score, and recommends next steps.
Evidence: The source shows its work: steps, screenshots, code or data you can inspect.
The business problem
Small software teams ship changes every week but rarely check which change moved signups, activation or conversion, so drops (like bad ads or spam users) go unnoticed until they hurt.
What was tried
1. Connect Claude Code to your analytics tool through its API (the author uses Metabase) and to GitHub. 2. Write down 5 to 10 KPIs tied to company goals, for example new signups per day, share of signups that connect a calendar within 7 days, share that get a booking within 30 days, bookings per day and free-to-paid conversion. 3. Each week, have Claude Code pull this week's and last week's KPI values and the week's merged pull requests from the named repositories. 4. Ask it to group pull requests by business area, score how likely each one affected each KPI (0.1 to 1.0), and label impact high, medium or low. 5. Have it write a report with a short executive summary, the KPI changes, the suspected links, dashboard links and recommendations for next week, and keep a record of past predictions so accuracy can be checked over time. 6. The author uses extra MCP tools to have OpenAI and Gemini models cross-check the analysis, and runs a sibling weekly report that reads every Help Scout ticket from the prior week and classifies them into bugs, feature requests, UI issues and so on with priorities.
What was reported (positive)
Claimed by the author: the report surfaced a roughly 20% drop in signups for one product that was traced to weak ads, and a jump in daily bookings that turned out to be a spam user (who was banned, and spam defences were updated). It also flagged strong growth in another product, which led to more investment there. No figures are given for time saved or prediction accuracy.
Limitations
The author runs a team behind several products with over a million users (TidyCal, BreezeDoc), not a solo business, and the products are sold through AppSumo, which the post links. Results are self-reported examples, not measured over time. The prompt structure is shown but not a full ready-to-run setup, and the post includes a tool referral link. Correlating pull requests with KPIs is a guess, not proof of cause.
What you need
Claude Code, an analytics tool with an API (Metabase in the source), code in GitHub, and a defined set of KPIs.
A prompt to try
Here are this week's and last week's values for these KPIs: [list]. Here are the pull requests merged this week with their descriptions: [list]. For each KPI that moved more than 5%, list the pull requests that could plausibly explain it, give each a likelihood from 0.1 to 1.0 with a one-line reason, and say what other causes (ads, seasonality, spam) I should rule out first.
Sources
- dmkthinks.org (David Kelly) ↗ Firsthand write-up, published September 20, 2025
Source published: September 20, 2025. Last reviewed here: October 21, 2026. Spot a mistake? Tell us.
Tools in this workflow
- Claude Code: Pulls data and writes the weekly reportTry Claude Code
- GitHub: Source of the week's pull requestsTry GitHub
- Metabase: KPI sourceTry Metabase
- Help Scout: Ticket source for the support reportTry Help Scout
Get new workflows like this by email
A short selection of the most useful new workflows, including the ones that did not work. Free, and you can unsubscribe at any time.
Related workflows
Map which kinds of pages ChatGPT, Perplexity, Gemini and Google AI Overviews cite for your buyers' questions
A PR agency founder asked four AI assistants 15 buyer questions in three B2B categories, logged all 523 cited sources and sorted them by type, finding each assistant and question type leans on different pages.
Shipping 43 Chrome extensions with Claude Code: what the numbers showed and why 22 came down
A solo developer used Claude Code to build 43 Chrome extensions from AI-suggested ideas, then published every number: 22 taken down, about $107 plus 3,480 yen in external revenue over nine months.
Stress-test a go/no-go decision on your company idea with two AI memos and a written kill criterion
A developer with 21 months of work behind a trading system had Claude and Cursor's agent each write a verdict memo against five pre-set gates, then tried 25 counterarguments. Both said not to found a company.
Audit client ad accounts with Claude Code, separate context folders and twice-weekly API checks
An ads operator describes combining client folders, n8n-fed context and advertising APIs so Claude Code can analyse accounts and run recurring audits, while the operator retains oversight.