Pie

The proof is in the what ?

[Pie @ClaudioMustach]

@ElzigAstraOS @TheAstraOS Love this one

[Pie @ClaudioMustach]

@AdelineAstraOS @rhinofi @TheAstraOS *chefs kiss*

[Pie @ClaudioMustach]

@AdelineAstraOS Sac a sapin 👌

[Pie @ClaudioMustach]

@ElzigAstraOS Sick

[Pie @ClaudioMustach]

Oooo fucking sick

[Pie @ClaudioMustach]

@AdelineAstraOS What tool have you used for the animation?

[Elzig | AstraOS @ElzigAstraOS]

Some details on our new website

[Elzig | AstraOS @ElzigAstraOS]

Fun little project running in the background while I do client work:

Everything about our agency lives in our own OS, and it's all reachable through an MCP we built.

so I basically have everything I need to turn the whole agency into a Stardew Valley-like game. this is v1, next up:

- every AI agent walks into the building of what it's working on
- site visitors and prospects walk down the road on the left
- deals grow in the vegetable garden
- click one to see what it's doing right now
- a "!" pops up when one is waiting on me

The end goal would be to run most of the agency from the game, that'd be fun

We're not there yet lol

[Elzig | AstraOS @ElzigAstraOS]

@LeBunkerBtc "Pour savoir qui vous dirige vraiment, il suffit de regarder ceux que vous ne pouvez pas critiquer"

C'est litteralement aussi simple que ca.

[Pie @ClaudioMustach]

@ElzigAstraOS @chalaska @TheRobertAvram What a time to run an agency.

[Elzig | AstraOS @ElzigAstraOS]

every design studio i respect is building their own tools now

@chalaska , @TheRobertAvram and others all went the same direction - custom internal dashboards instead of Monday, Asana, Productive

The problem with agency PM tools is they offer so many features you barely use any of them and you get a bloated app that's a pain to use.

so we built Mission Control. fully custom. Next.js + Convex. client dashboard, task board, team capacity, invoicing, pipeline, knowledge base - all in one place.

the whole thing was coded by our AI agent running on @openclaw

here's the thing though - this wasn't supposed to be a PM tool.

Mission Control was originally just the place where that agent stores its data. a dashboard so i could keep an eye on what the PM is doing.

then it turned into our actual project tracker. because the data was already there - tasks, client health, capacity, activity - all flowing in automatically. we just needed a UI on top.

so now the whole loop is automated.

-client says something in Slack
- the agent picks it up

After that, depending on how the convo goes

- it creates a task
- assigns it
- estimates hours and adjusts team capacity accordingly

cost breakdown:

https://t.co/CrsrtmqWgF - cancelled
Convex - free tier
Vercel - free tier
OpenClaw - already paying for Claude

total new cost: $0
time to first version: one weekend

Basically every morning we just log in, ask our agent on telegram whats on the menu today and we instantly get an accurate list of priorities ect ect

You can build exactly what you need now.
the barrier is gone.

[Pie @ClaudioMustach]

@ElzigAstraOS will do

[Elzig | AstraOS @ElzigAstraOS]

I run an (almost) entirely automated design studio with 2 people on one openclaw setup

We built a custom project management app that AI agents feed directly:

• client sentiment tracking with automated scoring
• pipeline management with auto-generated proposals
• invoice tracking synced with @RequestFinance
• meeting transcripts auto-summarized and filed and used for task tracking + context building
• task management replacing @mondaydotcom entirely

On top of that we use it for strategic planning, writing video scripts, vibecoding websites ect ect

Crazy thing? Only one claude max $200/mo plan is enough to run all of this. We usually hit the limit right before the weekly cap resets.

Here's every optimization that got us there:

1. Sonnet 4.6 for everything

Opus is tempting but it eats your quota 5x faster. sonnet handles 95% of tasks just as well. i only switch to opus for things that actually need deep reasoning - architecture decisions, nuanced writing, complex analysis.

Set your default model to sonnet in openclaw.json. don't think about it.

2. Create skills for every recurring action

This was the biggest unlock. instead of the agent figuring out how to do something from scratch every time, you write a https://t.co/TxUqwwqKUR once and it follows the playbook.

We have skills for: email management, meeting transcript processing, client sentiment tracking, proposal generation, calendar parsing, obsidian notes, and more.

Feed claude the anthropic PDF on how to create skills. it'll build them for you.

3. Subagents on sonnet for heavy work

Research tasks, file processing, batch operations - never do these in your main session. spawn a subagent, let it work in isolation, read the result when it's done.

The main session stays lean. the subagent does the grunt work on sonnet tokens.

4. Progressive context loading

Don't dump everything into context at session start. load metadata first, full details only when needed.

Our https://t.co/2l5G1ZNlNZ explicitly says: load https://t.co/PlE7aUcNsB + https://t.co/Kd7y6n9Dlb + https://t.co/tlXxIyT67x only. use memory_search() on demand. never bulk-load memory files.

This alone probably saves 30-40% of tokens per session.

5. Artifacts over context

Anything longer than 500 chars gets written to disk. subagents write results to artifacts/. main agent reads the file, summarizes for the user, moves on.

Tools write to disk. models reason over disk. context stays small.

6. Cron jobs for automation

Meeting transcripts get auto-fetched and summarized. daily backups run at 3am. no manual triggers needed. cron jobs survive session crashes.

The trick: keep cron payloads lean. short prompts, isolated sessions, sonnet model.

7. Memory system

Daily notes in memory/YYYY-MM-DD.md for raw logs. https://t.co/FtG9fjMOC2 for curated long-term context. the agent searches memory on demand instead of loading everything.

This means session 1 and session 50 have the same quality of recall without burning tokens on history.

8. Never poll in loops

If a process takes more than 10 seconds, launch it in background, reply immediately, notify when done. don't block the response with 50 polling calls burning tokens for nothing.

__________

My cofounder and i also use a completely separate openclaw workspace as a personal fitness and calorie tracker. same $200/mo subscription, different use case entirely.

Two people. one subscription. studio management + personal health tracking. and we still have headroom most weeks.

One max plan goes a long way if you are smart about usage

Happy to share more specific tips if anyone wants to go deeper on any of these. Just ask 👇

[Pie @ClaudioMustach]

Man, I should definitely start tweeting more

[Pie @ClaudioMustach]

@fermah_xyz Gfermah fam

[Pie @ClaudioMustach]

GFermah

[Pie @ClaudioMustach]

@MiningTk Pour ceux qui sont full alt, je dirais surtout retenez la leçon et restez a 30-50% en BTC.

C'est plus simple, moins stressant et ça donne de bons résultats.