AI tools monitor
Summary: A scheduled workflow that runs once a week, scans a curated list of sources (vendor blogs, release pages, AI tool tracking sites), identifies new tools and updates relevant to a target audience, writes a digest file to a known location, and pings the user on Slack with a short summary. Runs autonomously; the user reviews the digest when the notification arrives.
What makes this automation-level: The trigger is the schedule, not the user. The AI decides what to read, what's relevant for the target audience, what to include in the digest, and what to highlight in the Slack ping. The user shows up after the fact to read and act on whatever's worth acting on. This is the canonical "AI does the task, you review the output" shape.
Map
The recurring workflow is staying current on AI tooling without spending an hour reading every week. It decomposes into: identify the sources to monitor, check each source for changes since the last run, judge what's actually relevant for the target audience versus noise, summarise the relevant items into a digest, save the digest somewhere durable, and notify the user the new digest exists.
Every step is a good AI candidate: source scanning is mechanical, relevance judgement is pattern-matching against a defined audience, summarisation is structured rewriting, and the notification is a simple integration. The user is needed only to read the digest and decide what to do with it.
Test
Initial test was a single manual prompt: a list of source URLs and an instruction to summarise what's new and relevant for knowledge workers adopting AI. The first version produced too much detail — comprehensive but unreadable.
Adding two specific framings sharpened the output: "filter for what matters to a non-developer audience" and "lead with the three most useful items, then list the rest with one-line context each." The same prompt run a week later confirmed the format held up across different content. From there, scheduling was a small step.
Integrate
The implementation depends on the scheduling capability of the tool the user has chosen. The pattern is the same across tools — a saved prompt, a recurrence, allowlisted sources, a notification target — but the specifics differ.
With Anthropic Claude routines: the routine is configured at claude.ai/code/routines with the digest prompt, a GitHub repository as the workspace, and a custom cloud environment that allowlists the source domains plus the Slack API. Output gets committed to the repository each week; the routine posts a short summary to a Slack channel via a bot token stored in the environment.
With OpenAI ChatGPT scheduled tasks: the prompt is saved as a recurring task. Output stays inside the ChatGPT thread by default; the prompt itself can include instructions to email the digest to the user, or notification can rely on ChatGPT's native push notification. Source access depends on what's available to ChatGPT (web search and connectors).
With Google Gemini scheduled actions: the prompt is saved as a recurring action. The Workspace integration is the differentiator — the digest can be written directly into Google Docs and shared by Drive, or summarised into an email through Gmail, without separate connectors. Source access uses Gemini's built-in browsing.
In all three cases, the design choices that matter most are the same: the source list is bounded, the output goes somewhere the user actually checks, and the notification is the contract that ties the run to the human review.
Compound
The source list compounds. The digest catches some items that turn out to be useful and others that don't; over time, sources that consistently produce noise get dropped and sources that consistently surface useful items get prioritised. The digest format also evolves — the early version was longer than anyone read, and tightened over weeks as the user noticed which sections actually informed decisions.
A second compound mechanism is the user's mental model of what to look for. Reading the digest each week trains pattern recognition for what counts as a meaningful tool launch versus a marketing announcement. That recognition flows back into the prompt itself: the framing of "relevant for the target audience" gets sharper because the user has clearer language for what they actually want.
Related pages
- scheduled-automation — The full pattern this example sits inside, including the design considerations that make a workflow safe to schedule
- agent-use-case-evaluation — This workflow is a textbook fit for the five-point filter: bounded sources, recurring trigger, reviewable output, low single-run risk, and a notification that closes the human-review loop
- compound-engineering — The source list and prompt format both refine through use, illustrating two parallel compound mechanisms in a single workflow