Harsh realist 🖕
Anywhere the Army sends me
Joined September 2010
- Tweets1.1K
- Following60
- Followers128
- Likes94
Pinned Tweet
Add to your agent instructions when implementing with test driven development:
- Tautological tests considered harmful.
- Change-detector tests considered harmful.
- Do not create regression tests for bug fixes without a genuine gap in behavior testing.
pro tip: don't set up projects when you're running on hour 28 of no sleep. Set it up months ago. couldn't figure out why calls from server to db were slower than they should be. Refactored, parallelized, etc
today realized the code was out of washington dc, db in oregon
lmao
28
Needed a use for the Xenon edge on Omarchy since I’m spending 99% of my time in it own.
Used my pstack for t3code and had opus cook up a dashboard. Next up is swiping to new screens that will show me my t3code chats, my grokbot, etc.
3
946
If anyone was curious that uses @t3dotcodes for all their work, as long as the main thread is making spin-offs using `delegate_task` they'll have a 1-hour TTL since it's a separate `query()` call to the Agent SDK. They're tracked as children by `thread:delegate-task:[main-ID]`
Claude Code users, change this setting if you use subagents.
Claude has prompt caching with a 1-hour TTL, but that’s only for the main agent. Subagents get just 5 minutes by default.
If you use long-running subagents, I recommend increasing their cache TTL to 1 hour.
Just add this to ~/.claude/settings.json:
"subagentPromptCacheTtl": "1h"
Not sure why this isn’t the default.
62
The famous pstack for t3code @t3dotcodes . It aims for real Cursor parity. One install symlinks the skills into every connected Claude config, adds sticky /p3-mode hooks to Claude and Codex, and keeps subagents in role. Check-plan, worktree audit, etc
github.com/waffleflopper/p3-…
2
156
The goals is to have it work like pstack does with cursor. /p3-mode sustains between turns, compactions, and subagents. Check-plan and worktree scripts, better symlink install, a more. Full post is here:
p3-stack: @poteto's pstack, rebuilt to run natively in @theo's T3 Code. Built off the basic skill-only port by @uzairansar.
Goal is to get closer to how pstack actually behaves in Cursor, rather than just skills sitting in a folder.
one install.sh does the setup, symlinks every skill into .agents and reads t3's provider settings to link them into the config dir of every Claude instance you've connected.
/p3-mode now stays on, like a Cursor custom mode, by adding hooks to every Claude and Codex instance connected to t3code. Injects mode reminder each turn and tells agent to reread the skill after compaction. Codex hooks are auto-trusted. '/p3-mode off' ends it.
Workers stay in their role, too. A delegate_task child whose brief starts with a p3-agent.md gets marked as a p3 worker, and native subagents get the same brief through SubagentStart.
I've also caught it up with OG pstack's upstream:
- Pool-first /setup-p3: name your models and every role gets mapped to one of them (to avoid it picking stupid opencode models if you have it attached).
- check-plan + worktree-audit scripts
- pstacks guide ported over, page by page, with a slight tweak to the art (it's "theo-ized" 🙃)
github.com/waffleflopper/p3-…
54
p3-stack: @poteto's pstack, rebuilt to run natively in @theo's T3 Code. Built off the basic skill-only port by @uzairansar.
Goal is to get closer to how pstack actually behaves in Cursor, rather than just skills sitting in a folder.
one install.sh does the setup, symlinks every skill into .agents and reads t3's provider settings to link them into the config dir of every Claude instance you've connected.
/p3-mode now stays on, like a Cursor custom mode, by adding hooks to every Claude and Codex instance connected to t3code. Injects mode reminder each turn and tells agent to reread the skill after compaction. Codex hooks are auto-trusted. '/p3-mode off' ends it.
Workers stay in their role, too. A delegate_task child whose brief starts with a p3-agent.md gets marked as a p3 worker, and native subagents get the same brief through SubagentStart.
I've also caught it up with OG pstack's upstream:
- Pool-first /setup-p3: name your models and every role gets mapped to one of them (to avoid it picking stupid opencode models if you have it attached).
- check-plan + worktree-audit scripts
- pstacks guide ported over, page by page, with a slight tweak to the art (it's "theo-ized" 🙃)
github.com/waffleflopper/p3-…
Watched @mattpocockuk's talks and used his skills but still felt something wasn't "clicking" for me. After @theo did his video about smart/dumb zones and "wider" prompts, my AI use changed and started getting more done; still didn't "click". Watched @poteto talks, read articles and skills, and used her create-verification-skill to create a verification system feature by feature of my project.... and my productivity is on freaking hyper drive. The things I got from reading pstack skills and stealing a few concepts were the missing link for me.
Feeling pretty bullish about some productivity tool ideas I have for frustrating processes we have to sludge through in the military. Super stoked!
1
132
For those that don't want to read between the lines here. Taking the cut to 200/mo plan by 50% into account, using this Sol 6.1 that's 1/5th the price is a 150% increase in the total "usage" you could theoretically do compared to today using Astra.
Introducing 6.1 Sol, near Astra intelligence at one fifth of the price of Astra and 95% cache read discount. It is an absolute workhorse.
Combined with ultrafast for 8X speeded available today for Astra and coming soon for 6.1 Sol.
81
If the AI Bros that sensationalize after every model drop were ever correct, Anthropic and OpenAI would have been "cooked", "done for", and "will never recover from this" 50+ times by now.
It's amazing. Boldly wrong 100% of the time and still going like they have credibility
32
Last night and today Sol is feeling extra fucking slow. Opus 5.5 already felt faster but holy shit now it’s just tortoise and hare.
1
85
Made a skill to help clean up code bases that find themselves littered with bad tests like the above tries to prevent: github.com/waffleflopper/ai-…
20
4,914
Absolutely incredible.
We just shipped Quiver 2.0 in @paper
This is the best SVG generation possible, it's an incredible leap forward
• Much faster AND better quality
• Cleaner geometry
• New Telos model for detailed work
Next level vector generation -- available today in Paper
1
205
Starting to get nervous about this reset hitting. Been burning hard and down to 20% and it still hasn't hit yet. Sweatin hard!
51
Peak authenticity.
America is not a spectator sport
Apparently Logan and I going to Washington meeting with the Vice President, the Secretary of the Treasury, the Secretary of War, and other leaders of this country really pisses some people off
Good.
Because I’d rather be in the room asking questions, learning, challenging ideas, building relationships, and figuring out how I can help than sitting behind a screen telling everyone else what they should be doing
Twitter fingers have never built a country
Think about how insane the American Dream really is
Two brothers with a camera and a crazy amount of ambition can build businesses, compete on the biggest stages in the world, employ hundreds of people, invest in the future of this country and one day walk into Washington and sit across from the people helping lead it
That should inspire people not offend them
Washington doesn’t belong to commentators.
It doesn’t belong to celebrities.
It doesn’t belong to politicians.
It belongs to the American people
And whether people like me or hate me I’m going to keep showing up
Keep learning.
Keep building.
Keep investing in American companies.
Keep creating jobs.
Keep supporting American innovation.
Keep meeting the people doing the hard work.
Keep sticking my neck out.
And keep trying to leave this country stronger than I found it.
The next generation needs to believe that America is still a place where you can dream impossibly big, work your ass off, get knocked down, reinvent yourself, and eventually find yourself in rooms you never thought you’d enter
This country gave me opportunities beyond anything I could have imagined
I intend to give something back
You can spend your life commentating from the cheap seats or you can get on the field
I know where I’ll be
God bless the United States of America🇺🇸
577
Credit where credit is due, had never heard the term “tautological” before @mattpocockuk. I just tried to explain to the agent I didn’t want tests that essentially restated the implementation. His phrasing worked better and helped me make the others more succinct as well.
14
8,005