Digital Personal Assistants and OpenClaw

OpenClaw took the internet by storm - at least in developer and IT-adjacent circles. If you haven't come across it yet:

Fireship on YouTube | OpenClaw itself

I jumped on the bandwagon (who hasn't?) and experimented. My primary objective: set up a digital personal assistant via OpenClaw and Claude models. I've made notes.

Digital Personal Assistant

Personal assistants (PA): a second brain and a place to offload the frustrations of day-to-day inane choices. Replying to sales pitches, writing emails that are just process, reviewing the calendar and flagging gaps, keeping context you don't want to carry. I've used mine as a translation tool with generous context, too.

Especially when you want to be present for the things that actually matter (like your children, khm) - it's real leverage to automate the parts of your life that don't require your full attention.

I've set one up via OpenClaw, then migrated off into my local toolchain.

Streamlining the workflow - it's the future, near-certainly. Once the technology matures, it'll enter the mainstream. The future Siri or Alexa will be built on a similar architecture.

Let me walk you through what I've got:

#	Workflow	Trigger phrases
1	Gmail & Calendar	"check my email", "schedule a meeting with X", "read emails on 'contract foo' and schedule a time with agenda"
2	Email as Valerie	"email X as Valerie", "send X from you"
3	Parcel tracking	"check my parcels", "where is [tracking number]"
4	Research & analysis	"is this price fair?", "look up X", "check this invoice"
5	Contact pitches & outreach	"make a pitch for X", "send X something interesting"
6	Static landing pages	"make a page for X", "build a concept site"
7	Intern liaison	"what's <intern> asking", "prepare answer to <intern> question about X"
8	Document drafting	"draft me a message to X", "write a PR description"
9	Presentations	"make me a presentation", "build a deck for X", "prepare slides for X"

Non-exhaustive. All my usual steps are baked into memory, followed to the letter. I do my presentations a certain way — the robot boilerplates everything, ready for me to write the actual text. With email, I'll write it, but the fine details are handed off and handled.

There's a kink of human psychology this really leverages:

It's far easier to fix a draft than it is to start anew.

People often joke about the easiest way to get answers on StackOverflow: just answer it yourself, wrong, and watch people run in to correct you.

It does wonders for getting you moving through the errands of a day.

OpenClaw

Do you need OpenClaw to build yourself a digital PA?

Resounding no.

My setup migrated from OpenClaw to Nanobot, sidestepping Anthropic's anti-OpenClaw measures. I avoided them until it was time to resubscribe - they were a joke - and decided to re-evaluate before throwing more cash at it.

The experience with OpenClaw itself was uneven. Session state accumulates faster than you'd expect — I had a single session balloon to hundreds of messages over several days before catching it. Subagent reliability was inconsistent: tasks completing silently without doing anything, no error, no log. I was never sure if it even failed, no errors or warnings were visible.

The cost model is also wrong for the use case. Paying top-notch model pricing designed for coding sessions and active work never felt right without strict controls. And the more you control it, the less useful it is! In the end, the incentives just clashed far too much.

Anthropic has taken active steps to break OpenClaw (and assistant) compatibility, which meant never being sure if their laughable detection attempts were bugging my environment out, or if OpenClaw itself was broken, or if it's just the non-determinism of LLMs acting up. Lots of things could go wrong, for a thing that's supposed to self-direct.

There's also a piece of architecture I built on top of all this: A human-to-agent delegation model - the kind of thing that ages well regardless of what runtime is running it. Once I realized I can use that in an interactive session instead of having OpenClaw run it... the value just wasn't there. I'll write more on this. Watch the newsletter.

---

I skipped something deliberately: I gave the assistant a name (you might've noticed). That decision goes deeper than it sounds — next post.