9 Claude Cowork prompt-templates that run my 8-hour workday in 47 minutes of active supervision

Forty-seven minutes that's the median amount of time my hands are on the keyboard during what used to be a full workday.

The eight hours of actual task time still happens. It just happens in the background while nine Cowork prompt-templates do the work and ping me when they need a decision.

Each template is between 8 and 40 lines. each targets one specific task i used to do by hand every day or every week. Each has a measured before/after time from 30 days of running it side-by-side against the manual version.

Total weekly time pulled back across all nine: 34 hours.

That's not a projection — it's the delta between my logged hours in April (when i didn't have these) and May (when i did).

Below are all nine, verbatim. copy them as they are, or rewrite the bracketed slots to fit your stack. The three properties that make them work are at the end.

None of this is about Cowork being clever, it's about the prompts being shaped right.

The setup

I'd been using Cowork since the April 9 GA. By May I'd installed all 11 official plugins, hooked up the 12 connectors I actually needed, and built nine slash commands of my own. The connectors and plugins matter less than people think. The slash commands matter more.

A slash command in Cowork is just a saved prompt with structured inputs. You type /*morning-brief*, Cowork pops a small form with the fields the prompt needs, you fill them in, walk away. The prompt does the rest.

Most public Cowork content is about which plugin to install. That's the wrong layer. The plugin is the kitchen. The slash command is the recipe. You can buy every kitchen on the market and still cook nothing without recipes.

These nine are mine. They've been pruned from a larger set of about thirty I tried, twenty-one didn't survive. The nine that did share three properties I'll explain at the end. The twenty-one that didn't share two failure modes I'll explain too.

Numbers in each section: median time on the manual version (left), median time on the Cowork version (right). Both from 30 days of paired runs in May.

1. Daily intelligence briefing: 47 min -> 4 min

Runs every morning at 7:30. Pulls overnight Gmail, Slack DMs, calendar for the day, Polymarket positions, and three news sources I track. Outputs a single page with three sections: what needs my attention before 10am, what can wait, what's noise.

The TERMINATION line is doing real work here. Without it, the same prompt expanded to a 1,400-word document with executive overviews and "key takeaways" blocks. With it, the output is one printable page. The "do not summarize the summary" clause cut output length by 60% with no quality loss.

2. Competitive landscape scan: 3h -> 18 min

Runs on demand. Pulls product pages, pricing pages, recent blog posts, X presence, hiring posts, and recent funding for up to 8 named competitors. Cross-references against my own positioning notes (in a Drive folder I point at).

"Do not predict the future" saved 40 minutes per run. Cowork's default move is to add a strategic outlook section at the bottom of every report. The strategic outlook is always the same three observations dressed differently each time. Cutting it doesn't lose anything.

3. Email triage and draft replies: 90 min -> 11 min

Runs three times a day: 9am, 1pm, 5pm. Sorts inbox by who's waiting and how long, drafts a response for each one that needs it, leaves drafts in Gmail with a label so I can review and send in one batch.

Two things on this one. The "last 5 sent emails" voice-matching is the difference between a draft I can send and a draft I rewrite — without it, Cowork writes business-school English; with it, the drafts sound like me to the point where I sometimes forget which ones I touched. I tried 3 and 10 as alternatives; 3 loses my voice on rare subjects, 10 over-fits to old patterns. Five is the working compromise.

The other line, "do not generate replies for threads you've already drafted", came from week-one logs where Cowork was redrafting the same threads every morning. Three days of duplicate drafts piled up in Gmail before I caught it.

4. Meeting prep dossier: 30 min -> 3 min

Runs 2 hours before any external meeting. Pulls every past touchpoint with the people on the invite, the most recent five emails, any shared documents, their public profile updates since we last spoke, and three open questions I should bring.

"Drop it." Cowork wants to put everything in. A 4-page meeting prep is a polite distraction. The 1-page version gets read.

5. Weekly status report: 2h -> 7 min

Runs Friday at 4pm. Pulls Linear closed issues, Notion docs created this week, Slack channel digests, and calendar to reconstruct what I actually did. Generates the status doc people on my team and clients expect.

The "do not invent metrics" line is the single most important rule in this whole article. Cowork will fabricate a number when it thinks the audience expects one. I caught it generating "team velocity" estimates with no source three times in week one. Now it just writes "not tracked" and moves on.

6. Document review with Q&A: 90 min -> 9 min

Runs on any uploaded PDF or doc longer than 10 pages. Reads the whole thing, generates a structured Q&A on what matters, flags inconsistencies, surfaces anything that contradicts something I've written before.

The last TERMINATION line is the kicker. Cowork's default output for any document review is a linear walkthrough that's 30% of the length of the original. Useless. The structured Q&A format takes the same input and turns it into something I can use in 30 seconds.

7. Polymarket position audit: 45 min -> 3 min

Runs three times a day. Reads my open Polymarket positions, current market prices, and news from the last 12 hours that mentions anything related to active markets. Flags positions that need attention.

The reason "do not recommend new positions" is in the prompt is structural — this template is an audit, not a position generator. Cowork will absolutely try to suggest new trades if you let it. Mixing trade suggestions into an audit makes both worse. Keep them in separate slash commands.

8. Research deep-dive: 4h -> 28 min

Runs on any topic that needs a fully-cited research brief. Uses sub-agents in parallel, one per source category. Each sub-agent reports back, the coordinator synthesizes.

Sub-agents are the single biggest saving in the whole stack. 4 hours to 28 minutes is not a tuning win, it's five workers doing five things at once. The TERMINATION line keeps each worker from over-fetching its own category.

9. Content repurposing: 90 min -> 12 min

Takes one finished long-form piece and generates platform-specific adaptations: X thread, LinkedIn post, blog excerpt for newsletter, internal Slack share, email blurb for outreach.

"Never use the same opening line across channels" — the most violated rule across the discarded twenty-one. Without it, all five adaptations open with some version of "Three weeks ago I noticed..." The result reads like a content mill across your accounts. With it, each channel earns its own hook.

What the surviving 9 share

The thirty templates I tried looked roughly the same on the surface. The nine that worked share three properties. The twenty-one that didn't, violated at least one.

1. Explicit termination criterion. Every surviving template ends with TERMINATION:. The line names a condition Cowork can check by looking at the output: page count, field count, structural completeness. Not "be thorough" or "be complete." Something checkable. Without this, every prompt expanded to fill the full Cowork session block. With it, the median session length dropped from 2h 20m to 14 minutes.

2. Structured output shape, not free-form synthesis. Every surviving template specifies the output as named sections with specific contents. The model writes into the shape. It does not get to invent the shape. The discarded templates had outputs like "summarize and synthesize" - undefined containers that the model filled with whatever felt impressive.

3. Role definition in the first line. Every surviving template names a role: chief of staff, audit-only, coordinator, repurposer. The role limits what the model thinks it should produce. The discarded templates had no role line and produced outputs that tried to be everything at once.

The two failure modes of the discarded twenty-one:

• No clean stop. Without TERMINATION, the model spent the time block iterating. Output got longer and worse, not better. This was responsible for 14 of the 21 discards.
• Mission creep. Without role definition, an "email triage" prompt would also try to predict reply rates and recommend new contacts. An "audit" prompt would also try to recommend new positions. The model wants to be helpful in every direction — the role cuts it off. This was responsible for the other 7.

How the 34 hours add up

The math on weekly savings, in case you want to check it against your own usage.

Per-template savings, multiplied by how often I run each one:

Raw weekly delta is 72 hours. The actual figure I report is 34 hours because:

• Some templates overlap (triage and meeting prep both pull from Gmail; I don't count that input cost twice).
• I subtract the time I spend reviewing Cowork output and tuning prompts about 35 hours weekly across all nine.
• I don't count time during which I was doing something else productive anyway (sub-agent research running in the background while I take a meeting isn't free time, but it's not displacement either).

After all the subtractions: 34 hours pulled back per week. The raw 72 is the theoretical ceiling. The 34 is what I see in my logs as actual freed hours doing other work or not working.

How to use these

Drop each one into Cowork as a slash command via the Plugin Create flow. Setup takes about ten minutes if you've never made a slash command before.

The bracketed inputs need to point at your actual data sources before the prompt works. The structure of each prompt assumes the connectors are already authorized: Gmail, Slack, Google Drive, Linear, Notion, Polymarket. If a connector isn't authorized, Cowork will fail loudly on that step rather than silently skipping it. That's the behavior you want.

Run each one for a week before tuning. The templates as written are conservative. After a week of logs you'll see one or two places where the output is doing something you don't want, usually a section you skip every time, or a metric that's always blank. Cut those out of the prompt. Don't add more.

The 34 hours per week is the number after tuning. Week one for me was about 18 hours saved. Week three was 34. The tuning matters more than the original prompts.

*Bookmark — all 9 templates as a single file. Repost if you'd run any of these tomorrow.

Telegram for the full library + my pruned 21 (with notes on why each one failed): https://t.me/+_ZWrQN7GuDA3ZDEy*