This is a submission for the Notion MCP Challenge
What I Built
Every Monday standup, someone asks: "What did you work on last week?" An...
For further actions, you may consider blocking this person and/or reporting abuse
nice! you added the blog tone option.. did you like the output of those blogs? or tested with different models. I feel like claude will give you the best results.
yes bro i tied with some ollama models and groq support models and also gemini best result was given by gemini and also yes all tones were good especially the professional kind, but I use casual mostly
The decision to keep harvest and publish as pure function calls instead of agent reasoning is the kind of restraint that separates production-grade pipelines from demos. I've seen too many projects route everything through an LLM just because they can, then wonder why their pipeline is slow and nondeterministic.
Your Zod deduplication issue is one of those bugs that eats hours and teaches you something you'll never forget. The
pnpm.overridestrick is underrated — worth calling out for anyone building with Mastra or similar frameworks that have their own Zod dependency tree.Curious about one thing: have you considered adding a "diff from last week" view in the Notion planner? Something like highlighting repos that are new this week vs. recurring ones. That kind of week-over-week context would make the Monday standup recap even more useful — you'd see not just what you did, but how your focus shifted.
Alright got new feature idea, "adding a "diff from last week" view in the Notion planner? Something like highlighting repos that are new this week vs. recurring ones." Really loved and appreaciate this
Genius approach! I do something similar for marketing: git log -> delta with published posts -> content to tell. Build in public fueled by real commits is the best storytelling. The 3-agent pipeline idea is smart. How do you handle commit messages that are too cryptic for blog content?
Really clean architecture decision to keep LLM usage isolated to just the narration step. I've built a similar multi-agent content pipeline for cross-posting articles across Dev.to, Medium, and Hashnode, and I learned the same lesson the hard way — early versions routed everything through the LLM, including the actual API calls and formatting. Switching deterministic work to direct function calls cut execution time in half and eliminated an entire class of hallucination bugs.
The fallback chain for narration is smart too. Having a deterministic template-based fallback means you never end up with zero output, which matters a lot when these pipelines run on a cron with nobody watching.
Question: have you considered adding a "highlights" filter before narration? When I have a busy week, the raw data dump can be overwhelming for the LLM. Pre-filtering to the top 5-10 most significant changes (by lines changed, PR impact, etc.) tends to produce more focused and readable output.
The 3-agent split is really well thought out. I run a similar multi-step pipeline for a large programmatic SEO site — harvest data, generate content, deploy — and the biggest win was exactly what you described: keeping deterministic steps out of the LLM path. It cuts token costs dramatically and makes debugging so much easier when you know which stage failed.
The fallback chain for narration is smart too. We had a similar issue where our content generation model would occasionally produce malformed output, and having a deterministic fallback meant the pipeline never fully stalled.
Curious about the Notion Markdown Content API — that sounds like a huge improvement over constructing block arrays manually. Does it handle tables and code blocks well, or did you hit any edge cases there?
I love the architectural decisions and discipline.
What you're describing is constraint-based decomposition: each agent has one task, clear inputs and outputs, and no side effects from neighboring contexts. The LLM only appears when reasoning is genuinely needed. Everything else is handled through a function call. This greatly reduces the risk of drift and hallucinations.
I've been working on the content distribution layer for Trail (trail.venturanomadica.com). It uses Notion as the single source of truth. Claude fetches my long-form posts, adapts them for each platform, and schedules them via Buffer for multi-platform sharing. When one stage overlaps with another's responsibility, it becomes harder to spot failures and determine where decisions are made. The three-step process stays clear because the contracts between steps are explicit.
The hidden insight from your lessons learned: structured JSON output is 3-4 times slower than plain text with frontmatter parsing. That’s the kind of thing people only discover after building it the hard way. Worth emphasizing in the main discussion.
Solid work.
Using Mastra for agentic workflow and distributing tasks into 3 agents was good implementation. The agentic workflow architecture is simple horizontal one, waiting for input from previous agent.
Thank you
LLM only where it adds value is the right call — keeps the pipeline predictable and debuggable. Curious if the Gemini fallback produces a noticeably different tone from the primary path.
Cool project! The 'scratch your own itch' approach to building tools is the best kind of open source. What's been the most unexpected use case so far?
This resonates so much. I've been doing something similar for my solo project — using git log as the single source of truth for marketing content. Every feature, every bug fix, every refactor is a potential story.
My workflow: I maintain a
posts-log.mdfile that tracks what's been published. The delta between that file andgit logis literally my content backlog. No brainstorming sessions needed — the code writes the narrative.The 3-agent pipeline approach is clever. I've been doing it more manually (Claude Code session → draft → publish), but automating the GitHub-to-Notion step would save real time. Especially when you're shipping fast and forget what you built last Tuesday.
One thing I'd add: commit messages matter more than you'd think. Good commit hygiene = better auto-generated content.
Have you measured how much engagement the auto-generated posts get vs. manually written ones?
That is actually an interesting workflow. I have been trying to stay more consistent with writing, and using weekly GitHub activity as a base sounds practical. Do you usually just summarize commits, or do you expand them into more detailed write-ups?
Also curious if you are doing this manually or using some kind of automation.
Fun tool! I love it!