runesleo

If you use AI to track your to-dos, watch out for one thing: the list can grow too long for AI to read all at once.
I asked AI to keep all the things I need to handle personally in one file. Over time, the file kept growing until it exceeded what AI could read in one go. It never saw the part that didn’t fit, but still told me, “Everything’s taken care of.”
I’ve now changed two things:
· Move completed tasks to a separate “Completed” file, leaving only unfinished tasks on the list
· Set a size limit for the list and archive the oldest items before it gets too big
A list AI can’t read in full i

TapeOut ecosystem mining machines can now be split into shares for trading.
@Ignixbot launched mining machine fractionalization today. The first batch is @Fabx_global’s mining machines, and BEM mining machines are also on the way. I reviewed the rules and contracts and will explain them in plain English:
How it works
1. Deposit a mining machine into the pool, where its base hashrate is converted into $FABS shares
2. The mining machine continues mining as usual in the pool. 85% of the pool’s total output goes to FABS holders, 5% to liquidity providers, and 10% is charged as a fee
3. To get a m

Before bed, I now give AI a few tasks to work on overnight, so in the morning I only have to review the results.
Last night I assigned three tasks: review all the code written by different AIs during the day; cross-check my notes to see whether the decisions made during the day were recorded; and verify online, item by item, whether a batch of messages about pending tasks was true.
In the morning, I reviewed the results: the first two were ready to use, and the code review found 11 issues. The third task was rejected—it had marked the entire batch as “couldn’t find.” When I looked into why, I

Starting next week, Claude will default to “Auto Mode” on web, desktop, and mobile: it won’t ask you at every step anymore, and risky actions will be caught by an automated check.
I’ve been using Auto Mode for a while, but never fully trusted that check, so I added another layer: scripts hard-block actions involving money, posting, or account changes, and only let them through when I reply “send.”
Today I actually found a loophole: the quote repost tool wasn’t on the blocklist, so previously scheduled quotes relied entirely on the AI to wait for my approval. That’s fixed now, and Claude and Cu

Added to the Chinese Community Building group by the @BruceBlue and @TapeOutCF community foundations. Thanks 🙏!
A few things I built in the TapeOut ecosystem are all open source:
· tapeout-harvester: Automatically claims BEM rewards and stops if something goes wrong—no forcing through
· Agent Firewall: Runs a rule check before an AI wallet signs
· FabX Miner Desk: A read-only dashboard showing FabX miners their entry costs
I most agree with this line from the foundation’s post: whether the code runs and whether anyone uses it should count more than followers and holdings. Keep building~

Today alone, I used up 20% of Claude’s weekly quota, and I’m a little worried😅 At this rate, I’ll run out by Thursday.
Looking back, it wasn’t big tasks that ate up the quota, but having Claude do all the grunt work itself: writing scripts, scanning campaigns, changing configs—it did everything.
This afternoon, I changed the division of labor: Claude will just be the brain, responsible for breaking down goals, assigning tasks, and reviewing the work; each long-term workstream will be run by a Grok Bot lead, with one each for video, open source, and creator events.
Tomorrow, we’ll see whether

“One AI comes up with the plan, one writes, one checks”—I put it into practice all day today (Claude planned and reviewed; Cursor wrote the code). Here’s a workflow you can copy directly:
I. Make these 4 things clear when assigning tasks
1. Which files to modify and which ones must not be touched
2. The completion criteria: for example, “Show a preview of changes to real data first; do not write directly”
3. Tests that must be added, and tests must not touch real accounts or data
4. What to include when handing it back: what changed, test results, and a preview of the changes
II. During review

After reviewing the projects I've participated in, I noticed a painful pattern:
My unrealized gains at the highs were always solid; several projects reached multiples of their entry price.
But what I actually cashed out was nowhere near as much.
The problem wasn't picking projects; it was not having a selling plan in place ahead of time.
I've now set a few rules for myself:
1. Write a selling plan when I enter; take out my principal once it doubles
2. Sell in batches, don't gamble on the peak; selling too early is the cost of this strategy
3. Watch liquidity; if depth and volume drop, reduce m

@XCreators @Support My creator subscription price change has been stuck as "pending" since Sept 21. It's been over a month, and subscribers are still billed the old price.
I've opened a ticket and got a reply saying a specialist would follow up, but nothing has changed. Can someone take a look? 🙏

If you use Grok Bot and Claude at the same time, chances are half of the output never made it to your side.
I thought I could see everything the AI team worked on each day. Then I checked: I could read the conversations, but the actual results were in files on their cloud computers. When a cloud computer sat idle, it would disconnect from the internet. My side polled on a schedule, timed out every time, and didn’t even report an error.
The fix was to do the opposite: instead of me pulling the files, have the Bot use its connected computer to push them back. Memory syncs both ways, with the loc

Cursor’s quota pool works like this: Grok 4.7, Grok 4.6, Grok 4.5, and Composer 2.5 share one weekly pool, and Auto draws from the same allowance.
So there’s no need to use Auto day to day; just select Grok 4.7 and set the tier to Extra High.
My current workflow: once I run out of Claude quota, I switch to Cursor’s Grok 4.7.

I ran into another launchd pitfall.
I have an on-call AI that periodically calls codex exec to triage findings submitted by other AIs. Running it manually in the terminal works fine, but scheduling it with launchd means it doesn't do any work.
I found the cause: when launchd starts a process, its working directory is the root directory /, which isn't inside any git repository. Codex checks by default whether the current directory is a git repository and refuses to run if it isn't. The fix is just to add the --skip-git-repo-check argument. I also added a test while I was at it, so the default c

My ChatGPT Pro subscription runs through mid-November, and I’m not renewing for now.
Lately, the GPT models haven’t been very practical for my real work, and starting October 30, Pro’s Codex and Work quotas will also be cut in half.
Going forward, my main tools will be Claude for making judgments and wrapping things up, Grok Bot and Grok Build for monitoring X and researching, and Cursor for coding.
I’m not quitting for good. If ChatGPT releases a truly powerful new model or a major update worth subscribing for, I’ll come back. Subscriptions follow the models, not habits.

I have AI generate read-only daily reports for several X Chat groups. One item is to check whether there are any signals of rule changes, security risks, or similar issues in the groups. Yesterday it flagged a “rule/risk change.” I clicked in to check, and the original message was someone saying that a certain way of writing something was a bit hacky. The English risk terms were matched by substring, and hacky happened to contain hack, so it triggered. The English terms have now been changed to whole-word matching, so this false positive is gone.
The biggest problem with keyword alerts is fals