i didn't use openclaw correctly 🦞

update on some of the ai crap i’ve been up to

i got frustrated with the complexity i seemed to be having

as my understanding of ai increases, i found that wasn’t necessarily using them the right way

it’s not that i didn’t set it up right, (i don’t think), i’m sure people much better than me have something to say about that, but i think my usecases with openclaw were dubious at best, and i wanted to get this post out to correct any misinformation

i was using openclaw to:

1. manage my calendar and email (before i nuked it due to security reasons)

2. manage a video conversion, renaming, and uploading flow

3. create agents to manage my car service schedule and a ramit sethi like character to chat with about my personal finances

so we’ll work backwards on why i think all of these aren’t great use cases (for me)

i was building my ramit agent (and car agent) as another agent to be called on by openclaw, which doesn’t make sense

i’d say “activate car agent” or “ramit agent” and start giving it information for it to tabulate and keep track of

it was less obvious with car agent,but with ramit egent, even though i explicitly instructed my ramit bot to play ramit based on the knowledge and conversations given, he still responded with “ramit would say blah blah blah blah”

this was because my main orchestrator was acting as a middle man

i’d give some inputs, my openclaw would spin up ramit and talk to him — and then my orchestrator would relay the message to me and tell me what they (ramit and my orchestrator) talked about

this is a game of telephone — this is not what i wanted

openclaw is a harness that is meant to be an orchestrator, good at spinning up subagents when needed to help with tasks — like a good assistant should (depending on how powerful your model)

what i wanted was a main agent that played the role of ramit, not a subagent that reported to a main agent that reported to me 😬

so, i created a completely separate harness (harness being the runtime around a raw model (what tools it can call, what information it can access, how state and memory is managed, etc):

• podcast, book, interview extracts that become ramit’s voice

• access to the internet

• some kind of rolling memory window and access to summaries of past windows

• saved conversation state

• explicit not to dos, etc

i rebuilt that here1 completely separate from an orchestrator like openclaw

it’s working much better

if anyone wants to test, i can give you a telegram code, and you can start chatting with it2

i’m getting some good advice! 😂

however, everyone will burn tokens on my openrouter account since that’s hardcoded (for now) — perhaps that’s the next thing to abstract out

previously, i was using openclaw to help me move, convert, rename, and upload youtube videos based on very specific criteria

because i kept seeing inconsistent failures, i asked myself — this seems stupidly hard for a simple process, what’s the easy way to do it?

1. do it myself (not easy, but works 100% of the time, 60% of the time when I feel like doing it)

2. use a script (cron) — have it move, convert, and upload at specific intervals, and everything runs off code that can be verified and logged

openclaw was already just saying “oh ok frank wants this video upload stuff to be done, let’s pull up process.sh3 and run these commands”

ok loser, im just gonna automate this away from you because you’re just a redundant step in the process, especially if you’re telling me you’re just using a script

the pipeline required no creativity, just steps followed

so, i used claude code to shore up some cron jobs, and make sure the process works:

1. i airdrop to mac mini

2. 11pm it moves videos to the right folder

3. 3am starts the conversion and hard-drive saving process

4. 10am starts the upload process

5. after upload, the script cleans up and trashes done files

6. i come to check the work is done

7. rinse and repeat

no more “hey snappo, do this”

i also had to reach out to youtube to increase my video upload quota, otherwise i’d be stuck at 6 per day

next step, i’ll probably generalize this and publish it as a claude skill somehow: isolate files, convert & rename, upload to youtube 👍, but really, it’s a process that an agent or human can run

playing product manager on myself, i’d sit with myself and observe how i’m using openclaw to do scheduling

the honest truth was that the true delta between just telling claude code to put it on my schedule versus me going in and quickly adding a calendar slot was pretty much 0

and the error rate of me doing it myself was also 0

maybe there’s some kind of switching cost, but eh

this kind of outsourcing could be better for someone busier than me, but i pride myself on keeping a relatively minimal calendar schedule (and also email inbox)

inbox zero is not foreign to me and i manually unsubscribe to shit all the time

perhaps that’s a better use case instead — an ai that manually checks your email, gives you a list of shit that you could unsubscribe to, and you check yes/no and it goes off to do so

for me, as i check my email, i’ve been doing this manually, and over time, the effort has compounded so that my inbox is relatively uncomplicated

so uncomplicated inbox and calendar situation? doesn’t really necessitate the use of such a powerful, customized tool like openclaw, even though it was fun and fancy for a while