Every article I read about "AI agents for work" tells me the same things. Clean up my inbox. Summarize my meetings. Draft my emails. Turn rough notes into polished decks. Keep my calendar tidy.
These are all lovely ideas. They are also, almost entirely, a list of low-leverage work that I already don't need help with or don’t want to do better.
I've been using Claude Cowork for a while now. Here's what I've actually found useful, what I haven't, and the single most useful thing it did: telling me what to use it for (wait that bit was supposed to go at the end, oh well.)
I’ve also shared some of my scheduled tasks, in full, at the end of this post - copy and adapt them!
When I started with Cowork, I did the dutiful thing and read the "top use cases" posts. The advice was always a variation on the same list: inbox triage, meeting notes, polished drafts, calendar wrangling, create a PRD, etc.
The problem: I am already a fast reader and typer. I’ve been inbox zero since 2013. I’m now able to dictate with Wispr Flow at roughly the speed of thought (which is wrecking my WPM). The ambient chore of just reading stuff is kinda where I figure out what I think about things. Hand-typing or dictating is still faster and better for me than handing a draft off to an agent and shepherding the output back into something I'd put my name to.
So "help me reply to emails" turns out to be the equivalent of asking Cowork to help me pour coffee. Possible! But not obviously the highest-leverage deployment of a reasoning engine in the universe.
This is probably why I have historically sucked at the productivity-hack genre and have used Asana like a noob for 12 years.
The single most useful hour I spent with Cowork was asking it to help me figure out what to use Cowork for.
I started by complaining about the things I don’t want to use it for. I then gave it a bunch of context about me (more later) and what I’m already doing in a typical week, rather than what aspirationally I think I should be doing. The suggestions were much better than anything I'd have come up with on my own.
Two of my favorite scheduled tasks came directly out of that conversation.
Every Monday at 08:08 London time, Cowork pulls the quarterly objectives for every small team (about fifty of them) from the website repo on GitHub. Then it reads them all in one pass, looking for the stuff you only spot when you actually have every plan in your head at the same time: two teams using the word "activation" to mean different things, Team A's plan assuming Team B will ship something that Team B has not committed to, two teams pulling in opposite directions on the same metric, objectives still labeled Q1 in late April.
It posts a list to a private Slack channel, grouped by the person who owns each team. If nothing's worth flagging, it doesn't post at all. This work is the kind of cross-org program management that you basically pay executives fabulous amounts of money to do.
Also Monday, before I open my laptop, Cowork pulls current PostHog metrics via our own MCP, reads the growth review sheets from the sales and marketing teams, and scans the last 7 days of the main GTM channels in Slack. Then it generates 3–5 aggressive growth bets that I might not have considered.
Each bet has to name the “sacred cow” being sacrificed (CAC payback, margin, attribution clarity, brand consistency- pick one), rough size, a concrete first move with the person who'd own it, how we'd know to kill it, and a one-line rebuttal to the predictable frugal objection. The prompt literally says "hype man, not McKinsey" and "most spicy first."
I don't take all of the bets seriously. I take about one in five, which is roughly the hit rate I want, adapting as needed. I then copy and paste my favorite to the relevant team and pretend I came up with it myself.
I also periodically ask Cowork to check its own setup! “Where are you getting that from?” or “What context do you have on blah?”. This works surprisingly well - often Cowork will tell me about things like duplicate or stale files, and then suggest a solution.
The thing that actually makes all these suggestions work is the context the agent has to reason about your work. Ok this is boring and obvious, so I’m going to be extremely specific about what you actually do.
Out of the box, an agent knows nothing about you. It doesn't know Joe runs the Marketing small team (or even what a small team is), or that "activation" means something specific in our billing flow, or that this quarter the CSM north star is 120% NRR. Before you schedule anything, you have to hand over the context and have a way to keep that context fresh.
My setup lives under ~/Documents/Claude/Memory/:
CLAUDE.md is the index. Who I am, my direct reports (with GitHub handles and Slack IDs), current focus areas, standing preferences ("Meeting Free Tuesday — protect for deep work"), and a map pointing to everything else. glossary.md explains stuff to Cowork. Every PostHog-specific acronym (NRR, TAM, MCP), every Slack channel that matters. When someone pings me "TAM coverage proposal status?", the agent knows TAM = Technical Account Manager at PostHog, not total addressable market. It also knows that when I refer to James, I’m probably talking about James Hawkins not James Greenhill (sorry Jams). people/[name].md – one file per person I work closely with. What they own, how they want to be managed, what I've learned about working with them. This is the bit that helps Cowork help me prep for things like 1-1s. context/company.md, context/handbook/, context/exec-responsibilities.md – how the org actually works. Which exec owns which small teams. The current quarterly goals. How the handbook's structured and where to find what. (Side note – having a public handbook that consists entirely of .md files is awesome for this!)To keep the context fresh, I have a bunch of scheduled tasks running, which all started with a prompt like ‘ok so how are you keeping this up to date’:
CLAUDE.md. The issue is the only place I write my monthly priorities down for the whole company to see.The effect is that when I open Cowork on a random Tuesday morning and ask "who's the TAM for blah?", the agent already knows that it's a customer lookup, that Vitally is the source of truth, what our data model looks like etc. Without the context layer, it has to ask me clarifying questions every single time, or make random wrong assumptions.
The bulk of the value I get is from tasks that happen to me on a clock, not tasks I remember to ask for. My Monday morning stack runs itself before I've opened my laptop:
company-internal repo for new threads I might've missed, pulls RFCs my direct reports have authored in the last week, and flags any open issue or PR where I've been tagged but haven't responded. About 90 minutes of context-loading that now just shows up as a set of posts in a private Slack channel. I have still not figured out how to get Slack to send me an actual notification, so like a caveman I have a daily task in Asana to ‘check private Slack channel’.
Later in the week:
Lots of things! But I’m 80% sure the models will just keep getting better.
I’m sure all of this stuff will get better. I’ve gone from ‘what’s an MCP’ to ‘wtf why does your MCP suck’ in the space of 6 months, so I imagine that companies just need to catch up. There are a couple of SaaS tools we use that don’t have good/any MCP support yet that I would love to plug into Cowork, but we’re not there yet.
I'm excited about where this goes. The whole ‘ask the robot how to use it’ and then scheduling tasks taken together has taken me from messing about and creating slop to genuinely deriving value from AI tools. As PostHog grows, I’m already finding Cowork helps me keep on top of what’s going on without feeling like we need to impose more structure or project management to get things done.
I also think a lot of the early disappointment people have with agents is the "help me with my inbox" type stuff, where you’re getting the agent to do some low leverage task that is also hard to get right because the bar is "exactly how I would have done it, but faster." The bar is much easier to clear when the task is "do something I wasn't doing at all."
My advice, if you're trying Cowork:
Below are simplified versions of three of the scheduled tasks I mentioned above. The originals have a lot of PostHog-specific detail in them (team names, Slack channel IDs, direct report GitHub handles, our growth sheets); these versions have been trimmed to the pattern so you can adapt them to your own setup.
Drop each one into Cowork, and point the [bracketed] bits at your own stuff.
Frequency: Monday morning. Gets you: a terse list of places where team plans contradict each other or have gone stale.
Frequency: Monday morning. Gets you: 3–5 aggressive growth bets designed to push against your own defaults.
Frequency: Midweek, once things have had a few days to either get delivered or slip. Gets you: every unfulfilled promise other people have made to you in the last 7 days.
If the scheduled tasks are the output layer and the memory files are the context layer, the MCPs are the plumbing underneath that connects the agent to the rest of my stack. A few I'd actually recommend, in rough order of how often Cowork reaches for them on my behalf. Cowork calls these ‘Connectors’ to stop non-technical people like me from being scared.
The one that changes what's possible. When I ask "how's ARR looking this week?" Cowork runs actual SQL against our data warehouse, pulls live numbers, and hands them back.
I also use it for customer lookups. I route "who's the TAM for [customer]?" through the PostHog MCP because we keep Vitally synced into our warehouse. Every customer question Cowork answers starts with an execute-sql call against that data.
Every scheduled task that does real work touches this one. The coherence check fetches ~50 teams' objectives.mdx files in parallel. The 1:1 prep searches each direct report's RFC activity for the week. The internal digest surfaces threads I've been tagged in but haven't replied to. The monthly memory refresh reads my "Things Charles cares about" issue and updates CLAUDE.md from it.
If your company runs any real work on GitHub (a handbook repo, an RFC tracker, internal issues), this is the highest-leverage connector to install. Especially the search_issues and get_file_contents endpoints — most of the interesting automations are built on those two.
Not for sending things on my behalf - I don't want Cowork posting to public channels pretending to be me.
For reading stuff it's great: scanning the last 7 days of a sales or marketing channel, finding threads I'm mentioned in, reading out a DM conversation I need to pick back up, feeding the promise tracker.
Cowork posts its own scheduled-task output into a private Slack channel that only I can see. The private-channel pattern is nice - my Slack is already the place I check first thing in the morning, so putting Cowork's output there means I don't have to open a separate app to see what my agent's been up to overnight. I’m still not in the habit of regularly checking Cowork itself for the output.
If I had more meetings this would probably be a lot more useful. Works really nicely when I need it though!
For a task management tool I mostly use this in one narrow way: the Monday 1:1 prep task writes discussion topics directly into subtasks under a recurring "1:1 prep" task, one subtask per direct report. I suspect the Linear and Notion MCPs are equally capable.
Not bad at reading my calendar, occasionally used by scheduling-related tasks. I still do all my own calendar management by hand though.
Included for completeness. Reading files is fine - the push-harder ritual reads the sales and marketing growth review sheets every Monday. Anything involving Docs formatting or Drive organization is currently a mess. I tried getting Docs to actually update my 1-1 docs and it was comically inept.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。