@cormachayden_ did some research on lmnt... they just sell overpriced electrolyte packets, perhaps you could use the publicity to sell your own healthier packets? Probably wouldn't need to invest too much into ingredients to get started. I think they just use three basic ingredients... (they also have a class action against them for false advertising re maltodextrin.... kind of ironic). Best of luck to you sorry you are going through this but will be a learning experience. Just be more careful w/ publishing
@Joshua_WD edited the video looks like the x mute icon was covering the cat
cat stays in the corner for now, no buttons in that area, like a kind of customer service icon zone
lol i think? i made my first 16 bucks today (after over a year of vibe coding) and i think it may have been an accident since the subscription is for people who don't want to see ads in my free app (and I haven't started running ads..) maybe he felt bad for me, so it was like a tip lol
@istifadilkcesi are you saying when you respond to a spam rejection, it is better to provide info showing architecture/ feature differences rather than just screenshots?
🥳 Week 39’s Product of the Week goes to Post by FreeSurf!
A truly free, open‑source scheduler that lets you queue up posts across platforms – no hidden fees.
Dive in: https://t.co/4GzMMgQNDu
@un_contract
See below for curated tweets covering advice around SEO over the past week. I try to cut through the noise for people to isolate worthwhile advice. If you would like to subscribe to my newsletter, please check my bio.
"Traditional SEO still works"
@SEOKeval mentions that traditional SEO is still working for his clients ie no focus on PR or listicles. The basics are still working including optimizing keywords, but especially backlinks seem to be an underlying feature that helps websites standout irrespective of AI fixation.
Source: https://t.co/cmIfEpObM7
"How to automate SEO with Opus 5.5 (Full Course)"
While I debated listing this given that automating SEO is usually not a good idea, still there are innovative takes on the matter. He mentions Viktor (and perhaps is financially invested in the project), which can ingest data from several sources including Google Search console, DataforSEO and Ahrefs to drill down on SEO strategies, isolate what is doing well and isolate backlink strategies base on competitor information. Not a bad read not just for the agent capabilities but the different SEO tools one can use to get a birdseye view of how one is doing and strategy to implement from an agent.
Source: https://t.co/sbu5357X0Z
"Design doesn't fix SEO problems"
@therobertmay explains how a site redesign is not a solution to getting more customers ultimately. SEO is more important for driving customers because just because a site looks better doesn't mean it performs better if the fundamentals of SEO are incorporated.
Source: https://t.co/YvVAKZUqoJ
"SEO basics still important"
@bunbeau suggests to not focus too much on recent trends in engineering for AI overviews, but instead focus on basic things like whether your website pages are indexed and long-tail keywords.
Source: https://t.co/kpFhKQDf7J
For my newsletter following different topics for start ups including AI, SEO and Social Media Marketing, subscribe to my Feedfree Digest newsletter
*this was posted with FreeSurf's Post, a free social media poster with an API (X api usage paid at cost)
"SEO is dead in 2026."
People love repeating this narrative. "Google is broken, AI traffic ate search, don't bother."
They miss the entire point of distribution.
Here is why organic search is still the ultimate leverage leverage engine (and why I keep building for it):
When launching micro-SaaS or digital assets, fast feedback from organic search is your best signal.
Last Sunday, I put a brand new project online.
No domain authority. No backlinks. Zero marketing budget.
Today: First organic click directly from Google.
(thanks DataFast by @marclou)
1 single click.
Looks tiny on paper?
Look at the mechanics instead.
Most founders wait months to test indexing, architecture, and intent. They build giant SEO strategies before validating if Google even cares about their core offer.
That's the slow, painful way.
In a lean framework, that first click isn't just traffic. It's validation:
Indexation velocity is alive.
The site architecture works.
The intent signal is strong enough to bypass the noise.
If I tried to buy that intent via ads, I’d burn cash testing unproven keywords. If I relied purely on social posts, the traffic would die in 24 hours.
Over-analyzing algorithm updates before getting indexed is the ultimate founder trap.
You have two valid paths today:
1. The Passive Cynic Strategy:
Complain about GEO, AI Overviews, and decaying organic search. Spend weeks arguing on X while building nothing.
2. The Pure Velocity Strategy:
Ship clean architecture, push instantly, and capture long-tail intent before anyone else notices.
The fatal mistake is believing the doom-scrolling over execution.
Don't build complex content machines for unvalidated ideas.
Ship clean.
Let Google discover it.
When search volume picks up, optimize for scale.
Until then, speed > theories.
See below for curated tweets covering advice around social media marketing. I try to cut through the noise for people to isolate worthwhile advice. If you would like to subscribe to my newsletter, please check my bio.
"Youtube's automated deactivation problem"
@TierlistMan39 along with others on youtube are facing increasing challenges from opaque algorithms that will demonetize channels individuals have presumably done significant work for, forcing many to speak out about the problem. A warning about some of the difficulties creating a sustainable income from youtube.
Source: https://t.co/ZDOq2YCWr4
"Importance of a social presence"
@anulagarwal mentions the benefit of developing a social presence can change one's career trajectory and open doors even if some posts many people don't care about, it's still worth the long term effort.
Source: https://t.co/V1PhSNmXxV
"Are LinkedIn newsletters the easiest way to grow your list?"
@jayclouse discusses a case study of the Vojtech newsletter which relied on Linkedin's newsletter functionality to build his subscribers for his own list of subscribers in Kit. A key lesson being how social media distribution can help one find subscribers for newsletters.
Source: https://t.co/97kRLVwhhu
"Stay within legal lanes with youtube for success"
@robertoblake mentions youtube can still be a great place to grow an audience despite difficulties mentioned from other creators as long as you follow basic common sense legal rules like copyright
Source: https://t.co/u0scLsk8yp
"Platform burnout"
@Boldyboy1975 mentions that while on the surface, platforms look to reward originality rather than clipped or borrowed content or AI content, then reality is much more complex, with burned out creators, low or slow pay and lack of clarity.
Source: https://t.co/I35Y2D6CBC
For my newsletter following different topics for start ups including AI, SEO and Social Media Marketing, subscribe to my Feedfree Digest newsletter
*this was posted with FreeSurf's Post, a free social media poster with an API (X api usage paid at cost)
Platforms spent years begging for original content creators.
Now they’re finding out what that actually costs them — and what creators actually get in return.
The pitch was simple: make something yours. Don’t clip. Don’t stitch. Don’t farm. Film it, write it, own it. We’ll reward originality.
The fallout is that “original” became a slogan, not a paycheck.
What they sold:
• “Original content is king.”
• “We’re raising the bar for quality.”
• “Creators who make real work will get paid.”
• New funds, new “original content rewards,” higher Partner Program thresholds, crackdowns on AI slop and clip accounts.
What creators actually got:
• Median earnings still sitting around a few thousand dollars a year for a huge share of people doing this full-time. Nearly half still under $10K. https://t.co/wzAXBU3TAZ
• Money concentrated at the top. The top 10% take most of the brand spend. The middle gets squeezed between mega-stars and cheap UGC. https://t.co/m2GGobCbwW
• Platforms raising the gate. YouTube doubling watch-hour and Shorts thresholds. X killing old revenue share and replacing it with a narrower “original content” program. TikTok’s rewards system flagging actual original films as “unoriginal” until someone makes enough noise. https://t.co/zOQUo9dlUo
• Payouts that look generous in a keynote and tiny in a bank account. A million views on some short-form programs is still lunch money. Brand deals became the real income — which means you’re not a creator, you’re a freelance ad unit with extra steps. https://t.co/Oj9DPUzkR5
• Burnout and payment chaos as the background noise: delayed invoices, 120-day waits, algorithm changes that wipe a month of work overnight.
The joke is the platforms do want original work. They just want it cheap, abundant, and fully owned by the feed.
Original creators take the risk: time, gear, ideas, reputation. Platforms take the audience, the data, and most of the ad money. Then they change the rules and call it “quality.”
So the fallout isn’t that original content failed. It’s that the people making it finally noticed the contract.
They were invited in as talent. They’re treated as inventory.
That’s the gap. And it’s why a lot of the smartest original creators are quietly building email lists, products, memberships, and businesses that don’t vanish when a platform redefines “original” next quarter., I’m on holiday Back on Spaces next week, here for the laughs love you all xxx
See below for curated tweets covering advice around AI. I try to cut through the noise for people to isolate worthwhile advice. If you would like to subscribe to my newsletter, please check my bio.
"Break open-source AI tools for learning"
@LearnWithBrij recommends instead of trying to learn about AI from tutorials, try immersing yourself in openly available tools like OpenClaw, n8n and Ollama (as well as others) to try integrating them into your workflow, breaking them and learning how to be an AI engineer in the process.
Source: https://t.co/p7qiTpnY1h
"Try a suite of open-source tools to build better AI agents"
@damkina7 has an interesting write up showcasing lesser known open-source tools to fine-tune your own custom AI-agents or potential AI "employees" (tools include Storm, Agno and Cline)
Source: https://t.co/vfMX8S0M6d
"Try the open-source reverse-skill for better security with AI"
@bigaiguy His tweet showcases a rising security functionality that one can integrate with different AI agents to help arbitrage security problems, called reverse-skill. Worth taking a look perhaps for security work.
Source: https://t.co/7eFS7jUDvv
"Build agent failure write-ups"
@gadi_neelesh recommends making logs of agent failures when updating your codebase rather than focusing on agent benchmarks. Makes sense especially if agents are built on fixing patterns but don't have context of your codebase complexities. Give them a roadmap for the wrong way of doing things so they can learn from it and do better in the future.
Source: https://t.co/Gg4mNXRcGf
For my newsletter following different topics for start ups including AI, SEO and Social Media Marketing, subscribe to my Feedfree Digest newsletter
*this was posted with FreeSurf's Post, a free social media poster with an API (X api usage paid at cost)
I think every serious coding-agent user should start building a benchmark nobody else can download.
Your own failures.
Not SWE-bench.
Not HumanEval.
Not a leaderboard.
Every time an agent wastes 40 minutes, misunderstands your repo, loops on a tool, edits the wrong file, or confidently ships the wrong fix...
save the run.
Because that failure might be worth more than another 1,000 generic benchmark tasks.
Imagine this.
Your agent gets:
Fix checkout occasionally charging users twice.
It spends 25 minutes investigating Stripe.
Wrong direction.
Eventually you discover the actual bug was two workers claiming the same database job.
Today you fix it and move on.
I'd save the entire failed run:
failure-042
TASK
duplicate checkout charge
MODEL
Model A
RESULT
failed
WHERE IT WENT WRONG
step 17
BAD DECISION
focused on Stripe retry behavior
MISSED EVIDENCE
two workers processed the same job ID
TOOLS USED
23
TOKENS
41k
TIME
26 min
Now six weeks later a new model drops.
Everybody asks:
“Is it better for coding?”
You don't need X opinions.
Run it against your graveyard.
Can it solve failure-042?
Then:
failure-019
where the agent edited generated files.
failure-031
where it fixed the symptom instead of the race condition.
failure-055
where it kept increasing a timeout instead of finding shared state.
failure-071
where it changed an API that was supposed to remain compatible.
Now you have a benchmark that actually measures:
Will this model make MY work better?
That's already useful.
But I think you can go much further.
Don't only replay the entire task.
Replay from the exact moment things went wrong.
Example:
STEP 16
Observed:
two identical job IDs in worker logs
STEP 17
Agent chose:
investigate Stripe retries
Freeze everything before step 17.
Now swap only the decision-maker.
Ask 5 models:
What would you do next?
Maybe:
Model A
→ Stripe logs
Model B
→ worker concurrency
Model C
→ database transaction
Model D
→ frontend double-click
Model E
→ worker concurrency
Now you aren't benchmarking who can regenerate an entire app.
You're benchmarking a very specific capability:
who makes the right decision at the exact point my old agent failed?
That is way more interesting.
I'd call these decision checkpoints.
Every painful agent failure should produce one.
Over time you get:
agent-bench/
auth/
expired-session.json
redirect-loop.json
payments/
duplicate-charge.json
webhook-retry.json
database/
stale-read.json
transaction-race.json
testing/
flaky-port.json
shared-state.json
repo-navigation/
wrong-module.json
generated-file.json
Now whenever you change ANYTHING...
new model
new system prompt
new AGENTS.md
new tool
new MCP server
new reasoning level
new agent harness
new memory system
...run the same failure suite again.
And measure:
SOLVED 47 / 60
AVG TOOL CALLS
18 → 11
AVG TOKENS
31k → 19k
WRONG TURNS
14 → 6
REGRESSIONS
3
That last number matters.
Because an agent update can absolutely make some things better and other things worse.
Public benchmarks might say:
+8%
Meanwhile on YOUR repo:
auth debugging improved
but
database migrations got worse.
Without your own replay suite, you'd never know.
This gets even better if you freeze tool outputs.
Imagine the original run called:
Git
database
browser
API
logs
filesystem
Instead of hitting the live world again during evaluation...
record those results.
Then replay them.
Same evidence.
Different model.
Now you're testing the agent's DECISIONS instead of testing whether the outside world happened to behave differently today.
And for every new failure:
add it to the suite.
That's the compounding loop:
AGENT FAILS
↓
SAVE FAILURE
↓
FIND FIRST BAD DECISION
↓
CREATE CHECKPOINT
↓
ADD TO PRIVATE BENCHMARK
↓
TEST FUTURE MODELS / PROMPTS / TOOLS
↓
NEVER ACCEPT THE SAME REGRESSION SILENTLY
This changes how I think about agent failures.
A failed 40-minute session isn't automatically wasted.
It can become a permanent test case for every agent you use in the future.
And imagine doing this for a year.
You might end up with 300 tasks specifically shaped like the problems YOUR codebase produces.
At that point, when a new model launches, you don't need:
“bro is it better than Opus?”
Run:
my-agent-bench --model new-model
and get:
OLD MODEL
213 / 300 solved
$42.10
8h 17m
NEW MODEL
241 / 300 solved
$31.80
6h 04m
IMPROVED
repo navigation
race-condition debugging
frontend fixes
REGRESSED
SQL migrations
dependency upgrades
Now THAT is a model review.
And I'd keep another set:
golden successes.
Runs where the agent did something exceptionally well.
Because you don't only want future systems to avoid old failures.
You also don't want an upgrade to destroy behavior that already worked.
So your benchmark becomes:
FAILURE CORPUS
“Can you avoid mistakes we've already paid for?”
SUCCESS CORPUS
“Can you preserve workflows that already work?”
This also fixes one of the biggest problems with prompt tweaking.
People change AGENTS.md.
Agent feels better.
Maybe.
Change system prompt.
Feels better.
Install a new skill.
Feels better.
Switch model.
Feels WAY better.
Two weeks later something weird breaks.
Instead:
change anything
↓
run private replay suite
↓
compare against baseline
Then you can actually say:
This prompt change:
+9 solved cases
-2 solved cases
18% fewer tool calls
7% more tokens
auth improved
frontend unchanged
database regressed
Now prompt engineering starts looking like engineering.
Not vibes.
And this isn't purely theoretical anymore.
There are open-source tools appearing that record Claude Code / Codex agent runs, replay them step by step, fork execution from specific points, and diff how behavior changes.
Which makes me think the next important developer artifact after:
tests
benchmarks
logs
traces
might be:
agent trajectories.
Because once AI is doing hours of engineering work...
the sequence of decisions that produced the code becomes something worth testing too.
My rule would be:
Every expensive agent failure should become a cheap future regression test.
Don't just fix what the agent got wrong.
Save the moment it got confused.
Then make every future model prove it wouldn't make the same mistake.
Your best coding benchmark probably isn't on the internet.
It's hiding in all the agent runs that pissed you off.
I made a social media posting software app that's free because it doesn't seem necessary to charge for a posting software when the social media API's are free (my app is free except for X posts at cost).
Consider trying out the web app (API included).