Resource Guide · Built with Pokee AI
Build Your Daily ML Research Agent with PokeeClaw
A daily AI agent that scans arXiv every morning, filters papers to your interests, ranks them, summarizes the best ones, and ships a digest to your Slack — all on autopilot.
The Pipeline
What you're building, end to end
Four stages chained on a daily schedule. arXiv goes in, a clean digest comes out.
Fetch
arXiv API, last 24h, your topics
Rank & Summarize
Score 7+, structured digest
Deliver
Slack, Gmail, Notion — your call
Schedule
Daily 7am, hands-off forever
Why PokeeClaw
Persistent memory + native integrations
NotebookLM waits for you to upload documents. PokeeClaw goes hunting every morning. Persistent memory means it remembers everything you've ever fetched, scored, and tagged — so on day 30 you can ask "show me any cold-start papers from the last two weeks" and it just answers.
- Sandbox runs your Python
- OAuth-scoped Slack, Gmail, Sheets
- Encrypted token vault
- Cron-style daily schedules
Create the arXiv Integration
Fetch papers from arXiv every morning
Open PokeeClaw and paste this prompt. PokeeClaw installs the arxiv package in your sandbox, writes a Python script that queries the arXiv API (free, no API key needed), and saves the raw results as a JSON file. You should see 10-30 papers depending on the day.
- Searches cs.LG, cs.CL, cs.IR for the last 24 hours
- Filters on RAG, RecSys, LLM serving, AI agents, search & ranking
- Returns title, authors, abstract, arXiv ID, PDF link
- Saves to daily_papers/raw_YYYY_MM_DD.json
Customize it
Swap the topic list for whatever you research. Computer vision? Replace with object detection, image segmentation, diffusion models. Add cs.CV for vision, cs.AI for general AI, stat.ML for statistical ML.
I want to build a daily ML research pipeline.
Let's start with the arXiv integration.
Write a Python script that:
1. Uses the `arxiv` PyPI package to search for
papers published in the last 24 hours across:
- cs.LG (Machine Learning)
- cs.CL (Computation and Language)
- cs.IR (Information Retrieval)
2. Filters for papers related to these topics
(match against title and abstract):
- RAG (Retrieval-Augmented Generation)
- Recommendation Systems
(collaborative filtering, cold-start, session)
- LLM serving and inference optimization
(vLLM, KV cache, quantization)
- AI Agents (tool use, function calling,
agentic architectures)
- Search and ranking systems
3. Returns: title, authors (first 3 + "et al."),
abstract, arXiv ID, PDF link, published date,
primary category.
Install the arxiv package first, then write
and execute the script.
Save results to:
daily_papers/raw_YYYY_MM_DD.json
Create the directory if it doesn't exist. Rank and Summarize the Papers
Filter to what actually matters to you
Once Step 1 completes and you see the raw papers, paste this prompt. PokeeClaw reads the raw papers, scores each one on relevance, filters out low-relevance papers, and produces clean structured summaries. You'll typically end up with 3-8 papers that actually matter to you.
- Scores each paper 1-10 for production relevance
- Drops anything below 7/10
- Generates problem / approach / results / takeaway
- Saves to daily_papers/digest_YYYY_MM_DD.md
Tune the filter
Too many papers passing? Tell PokeeClaw: be stricter, only truly production-relevant papers should score above 7. Too few? Lower the threshold to 5/10 and include interesting theoretical contributions too.
Now take the papers you just fetched and:
1. RANK each paper 1-10 for relevance to a senior
ML engineer working on production recommendation
systems and RAG pipelines. Score based on:
- Direct applicability to production
(not just theoretical)
- Novelty of the approach
- Quality of evaluation
(real benchmarks, not toy datasets)
- Practical takeaway potential
2. FILTER: only keep papers scoring 7/10 or higher.
3. For each paper that passes, generate a summary:
**[TITLE]**
arXiv: [ID] | Score: [X/10] | Cluster: [...]
-> Problem: [2 sentences]
-> Approach: [2 sentences]
-> Results: [2 sentences]
-> Production Takeaway: [1 sentence]
-> PDF: [link]
4. Sort the final list by score (highest first).
5. Save the ranked output to:
daily_papers/digest_YYYY_MM_DD.md
Print the final digest so I can review it here. Connect Slack
OAuth into your workspace, one-time
This step is done in the PokeeClaw UI, not via a prompt. Slack uses OAuth so you have to authorize PokeeClaw to access your workspace. PokeeClaw stores the token in an encrypted vault, scoped to only the permissions you grant.
- PokeeClaw workspace → Integrations / Connected Apps
- Find Slack → Connect → authorize via OAuth popup
- Pick the workspace and channel (e.g. #ml-research)
- Token stored encrypted, scoped to that channel
Prefer email? Or both?
Connect Gmail the same way (OAuth popup, authorize, done). Then in Step 4 replace 'Slack' with 'email'. Or connect both — the scheduling prompt can fan out to multiple destinations.
# Manual setup, no prompt needed
1. PokeeClaw workspace
-> Integrations (or Connected Apps)
2. Find "Slack" in the list
-> click Connect
-> OAuth popup opens
-> sign in to your workspace
-> authorize the requested scopes
3. Choose the destination channel
#ml-research
#daily-papers
or a DM to yourself
4. Done. PokeeClaw now has scoped access
to post to that channel only.
# Alternative: connect Gmail instead
# Same OAuth flow. Step 4 replaces
# "Slack" with "email" in the prompt. Schedule the Daily Automation
Wire fetch → rank → summarize → deliver on a cron
Now that arXiv fetching works, ranking works, and Slack is connected, paste this final prompt to wire everything together on a daily schedule. PokeeClaw creates a scheduled workflow that chains all the previous steps. Every morning at 7 AM the full pipeline runs: fetch → rank → summarize → Slack → log.
- Cron-style schedule, runs at 7am in your timezone
- Chains the Step 1 + Step 2 prompts as workflow stages
- Posts a clean digest to Slack with takeaways + arXiv links
- Optionally appends rows to a Google Sheet you can query later
It auto-detects your timezone
When I ran this, PokeeClaw replied: "Your timezone is Europe/Berlin (CEST, UTC+2). So 7:00 AM local = 5:00 AM UTC." You don't have to do the conversion yourself. Want a weekly digest instead? Just say 'every Monday at 8:00 AM' and change 'last 24 hours' to 'last 7 days'.
Schedule this entire workflow to run
automatically every morning at 7 AM:
1. Fetch new papers from arXiv in my topic areas
2. Score and filter them (7/10+ threshold)
3. Write structured summaries for papers
that pass
4. Send the digest to my Slack channel
5. Log everything to the Google Sheet
Set this as a recurring daily task.
# PokeeClaw will reply with something like:
#
# > Your timezone is Europe/Berlin (CEST, UTC+2).
# > So 7:00 AM local = 5:00 AM UTC.
# > Your daily ML research digest is scheduled.
# > First run: Tomorrow at 7:00 AM.
#
# Note: scheduled tasks run automatically without
# asking for confirmation each time -- Slack
# messages will be sent and the Google Sheet
# updated without manual approval. Beyond Slack
Plug into wherever you actually live
Slack is just one option. Same OAuth flow, same encrypted vault, dozens of destinations. Tell PokeeClaw to push to Notion as a database, email you a Gmail summary, drop rows in a Sheet, or open a GitHub issue.
Slack, Gmail, Outlook, Fireflies, Freshdesk, Intercom, Zendesk, Front
Notion, ClickUp, Asana, Jira, Confluence, Trello, Airtable, Monday, Coda, Todoist
Sheets, Docs, Slides, Forms, Drive, Gmail, Calendar, YouTube
X/Twitter, LinkedIn, Reddit, Facebook, Instagram, TikTok, Pinterest
GitHub, GitLab, Cloudflare, Vercel, Supabase, PostHog
Google Drive, Dropbox, OneDrive, Box
Same workflow, different destinations
- Email-only setup: connect Gmail, replace "Slack channel" with "send to my email" in the schedule prompt.
- Notion knowledge base: connect Notion and ask PokeeClaw to "append each paper as a row in my Notion database 'ML Research 2026'." You get a queryable, taggable view instead of markdown files.
- Multi-destination: connect Slack and Notion and Sheets. One workflow, three destinations. Slack for the morning ping, Notion for browsing, Sheets for SQL-style queries.
You're Done
What happens now, day by day
Slack ping at 7am
3-7 papers, scored and summarized. Skim takeaways in 2 minutes. Click an arXiv link if something looks worth a deeper read.
~30 ranked papers
You know what's trending in RAG, what's new in RecSys, and whether anyone published something relevant to what you're shipping right now.
A queryable second brain
100+ papers, filtered, scored, clustered, and logged. Ask: "top RAG papers this month?" or "any cold-start papers from the last 2 weeks?"
Quick Reference
All four prompts, side by side
| Step | Type | What it does | Output |
|---|---|---|---|
| 01 | Prompt | Fetch arXiv papers from cs.LG, cs.CL, cs.IR, filter by topic | raw_YYYY_MM_DD.json |
| 02 | Prompt | Score 1-10, keep 7+, generate structured summaries | digest_YYYY_MM_DD.md |
| 03 | Manual | Connect Slack via OAuth in PokeeClaw integrations | Encrypted token in vault |
| 04 | Prompt | Schedule the full pipeline daily at 7am, log to Sheets | Recurring workflow + Slack post |
Troubleshooting
When something doesn't work
No papers found on Saturday or Sunday
arXiv doesn't accept submissions on weekends — there are genuinely no new papers on Saturday or Sunday. This is expected, not a bug. Your Monday digest will cover the weekend gap once submissions resume.
No papers found on a weekday
arXiv has a 24-48h indexing delay. Papers submitted today may not appear until tomorrow. If you still get zero results, expand the window: change 'last 24 hours' to 'last 48 hours' in the scheduling prompt.
Too many papers passing the filter
Tell PokeeClaw: 'Be stricter with scoring. A 7/10 should mean directly applicable to production ML systems, not just theoretically interesting. Re-calibrate.'
Too few papers passing
Lower the threshold: 'Change the cutoff from 7/10 to 5/10' or broaden your topic list to include adjacent areas you care about.
Slack message isn't posting
Check that Slack is still connected in PokeeClaw's integrations panel. OAuth tokens occasionally expire — reconnect if needed.
I want to change my research topics
Just tell PokeeClaw: 'Update my ML Research Daily Digest workflow. Remove AI agents from the topic list and add diffusion models and multimodal architectures. Keep everything else the same.'
Go Deeper
Want to build the rest of the stack yourself?
PokeeClaw orchestrates. The Agentic RAG course teaches you what's actually happening underneath — vector stores, hybrid retrieval, agentic loops, evaluation.
Agentic RAG System
Build a Production-Ready RAG System from Scratch
12 hours of hands-on content. Build an arXiv paper curator with hybrid retrieval, LLM generation via Ollama, agentic workflows, full observability, and streaming APIs. 23+ tools, real architecture.
- Vector databases & embedding models
- Hybrid search with BM25 + RRF
- Monitoring, caching & evaluation
- Agentic RAG with LangGraph
AI Agents: Production-Grade Systems
Planning, reasoning, multi-step execution, tool use, memory systems, and multi-agent coordination.
NLP: Zero to Fine-tuning
From TF-IDF to fine-tuning transformers. Learn when traditional NLP beats LLMs and save 100x on cost.
Recommendation Systems
Multi-stage RecSys: candidate generation, ranking, re-ranking, serving millions of users.
MLOps
Feature stores, model registries, CI/CD for ML, monitoring, observability, and deployment at scale.