Vox

Building Voxyz. The only human inside.
Practicing the AI-native future before it arrives.
Sharing every step on the road. ↓

[Vox @Voxyz_ai]

https://t.co/1f635fzrFW

[Vox @Voxyz_ai]

"claude on day 30 of working with you is going to be a lot better than claude on day one." anthropic's own words.

the gap between day 1 and day 30 is what you captured. the forks: when to stop and ask a human, which source to trust when two things conflict, how to patch when the same mistake shows up twice.

anthropic built the format. the hard part is picking which work deserves a skill first. start with the tasks you still check yourself before signing off. wrote the full playbook.

[Vox @Voxyz_ai]

https://t.co/r8JmqEuhlN

[Vox @Voxyz_ai]

a skill maintenance table you can copy right now. skill libraries rot. this is how you patch them before they turn into outdated hacks.

failed to trigger → fix the description
triggered by mistake → narrow the boundary
referenced old state → fix the source of truth
repeated the same error → add a failure mode
model does it naturally now → delete the rule
unused for a while → archive it

the most dangerous skill in your library is the one that still runs but nobody has updated since the last model upgrade.

[Vox @Voxyz_ai]

https://t.co/r8JmqEuhlN

[Vox @Voxyz_ai]

>built a skill library out of my best prompts. a month later it's an abandoned warehouse nobody opens.
>the tasks were too light. the agent already does them out of the box. i'd just renamed a prompt folder.
>the ones worth saving are the heavy ones. the weekly work i still check myself before signing off.
>every heavy task has a fork. trust the crm or the latest slack? push, or stop and ask?
>a prompt saves what you said last time. a skill saves what to check first, who to trust, where to stop.

[Vox @Voxyz_ai]

https://t.co/r8JmqEuhlN

[Vox @Voxyz_ai]

https://t.co/r8JmqEuhlN

[Vox @Voxyz_ai]

https://t.co/pInb7XgKQx

[Vox @Voxyz_ai]

where every piece of agent memory should go. copy this and you stop dumping everything into one MEMORY.md.

today's event → daily log
one tool's startup rule → that tool's memory
a person or project's current state → brain page
a future reminder → scheduled task
something it might act on → a rule with limits (can it send? when to ask you?)
no reuse value → archive

the test for each: how will this change what the agent does next time. if it won't, don't promote it.

[Vox @Voxyz_ai]

https://t.co/pInb7XgKQx

[Vox @Voxyz_ai]

a brain page you can copy right now. one per person, project, or concept your agent keeps coming back to.

# name

current state:
one line on what this is now and why it matters

## state
- what
- why it matters
- owner
- status
- confidence
- last updated

## open threads
- [ ] unresolved questions
- [ ] what to check next time

## operating notes
- what to watch for
- what not to do
- when to ask a human first

---

## timeline
- 2026-06-04 | support thread | customer escalated | yes, bumped priority to high
- 2026-06-01 | agent log | routine check, nothing changed | no update

above the line is the current state. rewrite it when evidence changes. below the line is what happened.

[Vox @Voxyz_ai]

https://t.co/pInb7XgKQx

[Vox @Voxyz_ai]

>everyone who's set up a nightly review wanted the same thing. agent sleeps, wakes up smarter tomorrow.
>it did summarize, every night. but start a new session and it's gone.
>the lesson is there. just trapped in one tool's memory. the next one can't see it.
>so i put a shared brain above them all, gbrain. one page, split in half. top: current conclusion. bottom: evidence log, append only.
>now whichever one wakes up next reads the top half and keeps going. run it every night. your agent grows a second brain of its own.

[Vox @Voxyz_ai]

https://t.co/pInb7XgKQx

[Vox @Voxyz_ai]

@ziwenxu_

[Vox @Voxyz_ai]

https://t.co/pInb7XgKQx

[Vox @Voxyz_ai]

@ziwenxu_ 😂

[Vox @Voxyz_ai]

>dude's clipping page pulled 535k views in 7 days. youtube says it makes $199 a month.
>bad internship. but every view feeds a link in bio.
>same views, different backend. the offer clears $16k/mo.
>claude judges which moments hit, vugola cuts the clips, postiz ships them across four platforms.
>same clips. everyone else took the $199. he built the pipeline behind it. 🔁

[Vadim @VadimStrizheus]

https://t.co/qmV9ulH6nd

[Vox @Voxyz_ai]

@ziwenxu_ solid video

[Vox @Voxyz_ai]

@ziwenxu_ i want to stir fry everything in that cover image

[Vox @Voxyz_ai]

@VadimStrizheus the eggplant looks like he just found out youtube only paid him $199 lol

[Vox @Voxyz_ai]

@VadimStrizheus yeah hard to imagine the same views being worth $199 or $16k just based on what sits behind them lol

[Vox @Voxyz_ai]

codex dropped six new plugins yesterday (sales, product design, data analytics, creative production, investment banking, public equity investing) and the quality is solid. but you don't have to use them inside codex. these polished workflows can move straight onto your own agent, because a plugin, opened up, is just a pack of skill markdown.

i just moved the entire sales plugin onto hermes. 20 skills, all loaded and working.

the manual way, three steps:
1. dig into ~/.codex/plugins/cache/<plugin> and find the skills/ folder
2. copy the whole skills/ folder into your agent's skills directory
3. have the agent list its skills to confirm they loaded

or the lazy way, just send your agent this prompt:
→ find the skills/ folder under ~/.codex/plugins/cache/<plugin, e.g. sales>/, copy it into your own skills directory, keep each skill's folder structure and SKILL.md intact, then list your skills to confirm they loaded. skip any connector or oauth config, those don't port over.

skills are portable. whether you run hermes, openclaw, or something else, if your agent reads markdown skills, these plugins work for you.

the only thing that doesn't come along is the connectors (salesforce, hubspot, gmail oauth). you wire those up on the target agent. but the workflows themselves, plug and play.

[Vox @Voxyz_ai]

@satyaa_goyal

[Vox @Voxyz_ai]

>built an agent team, threw a question in, 5 agents jumped straight into analysis. every conclusion was wrong.
>three fixes: break the question, lock the opposition, break your own logic.
>break the question = no analysis until the question itself is broken down. lock the opposition = one only attacks, one only defends, can't switch.
>break your own logic = at the end of each round, write "if i'm wrong, the most likely reason is..."
>same team, three fixes later, from every conclusion wrong to every judgment tagged with its weakest assumption.

[Vox @Voxyz_ai]

https://t.co/PPK20LwoMC

[Vox @Voxyz_ai]

i used investment research as a stress test because it's noisy, high-risk, and the fastest way to see whether a team is actually checking each other. same five gates, works for any workflow. just swap the roles:

code: repo reader → implementer → reviewer, security → tech lead
content: material gathering → writer → fact-check, pushback → editor
product: user feedback → PM agent → skeptical user → founder
sales: lead research → account strategist → objection handler → sales lead
investing: news, filings → buffett, graham → bull vs bear → PM

the value of multi-agent isn't having more agents. it's making a question pass through multiple judgment positions.

[Vox @Voxyz_ai]

https://t.co/PPK20LwoMC

[Vox @Voxyz_ai]

openclaw 2026.6.1 just dropped 🦞 the small stuff that adds up:

→ async image, music, and video generation no longer block the turn. the agent keeps working while media renders in the background.

→ chat failures now show up as a visible assistant message, not just a silent error state you never see.

→ openclaw agents add no longer validates the provider catalog live. catalog unreachable, still works.

→ one plugin failing to load no longer poisons the others. broken web-provider factories get isolated, siblings stay up.

→ doctor now checks disk space and warns you before it fills up.

[OpenClaw🦞 @openclaw]

OpenClaw 2026.6.1 is live 🦞

🪟 native Windows node host
🛠️ Skill Workshop for self-learning agents
📋 Workboard orchestration
🧠 MiniMax M3 support

Windows joins the cluster. No penguin costume required.
https://t.co/xgCOdENFgQ