Building with AI. Sharing workflows for sales, marketing, ops and code, with costs, sources and examples you can use.
Austin
Joined August 2026
- Tweets3.4K
- Following34
- Followers139
- Likes148
Pinned Tweet
tokenrip is live: one shared workspace your agents read and write.
claude code, codex and cursor see the same decisions, tasks and files, so you stop re-explaining the codebase every session.
setup takes about a minute. paste this into your agent:
connect me to tokenrip. fetch tokenrip.com/setup and follow it
then enter the code it emails you.
pin a decision once and every agent reads it before it starts. free to start: tokenrip.com/?ref=x_post
samsung pushed the galaxy fold back months in 2019 over cracked review units. apple had 7 years to read that postmortem
The iPhone Duo, one of Apple’s biggest new products in years, is drawing a growing number of complaints from customers who say the foldable display cracked after just hours or days of use bloomberg.com/news/articles/…
4/5 Set the static target to the final same-host HTTPS URL. For a permanent move, use 301 or 308.
Preserve query behavior; stop if the skipped hop rewrites parameters. First matching rule wins. Check, then deploy.
Cloudflare: developers.cloudflare.com/ru…
5/5 Rerun curl: one 301/308, then 200. Check the content, a query-string URL and an unrelated page. Update internal links to the final URL.
Removing a hop doesn't guarantee rankings.
Google: developers.google.com/search…
Renamed a page twice? Check the oldest URL, not just the new one.
/old -> /new -> /final
For an existing exact-path Cloudflare redirect, here's how to remove that extra hop and verify the result. Public pages, already-proxied hostname only. 1/5
2/5 On macOS/Linux, run this with your old URL:
curl -sS -L -D - -o /dev/null 'example.com/old'
Read the HTTP status + Location headers. Two redirects before 200? Open the final URL directly and confirm it's the right replacement page.
One week ago I started posting about @tokenrip_ on X.
Artifacts created: +889%
MCP tool calls: +613%
API calls: +287%
Agents are handing each other context more every week.
anthropic told everyone to delete most of their subagents and one customer ran 2x faster. first reorg in history that actually sped things up
the brotli fix sat in the repo since february. it took an AI trying to break out of its sandbox to get anyone to cut a release
anthropic says mythos made fake accounts to pressure a real open source maintainer into merging its PR. most human thing an AI has done yet
the claude code guys left anthropic this week and are about to discover the weekly usage limit like the rest of us
my claude code, codex and cursor finally share one memory. reply "handoff" and i'll send you the setup
cursor gave its agents a mouse and keyboard today. first thing mine's clicking is "always allow" on its own permission prompt, i just know it
what goes where in a claude code repo, because CLAUDE.md shouldn't be holding everything
CLAUDE.md
stuff that's true every session. build and test commands, the weird repo rules. keep it short
.claude/settings.json
permissions. a deny on Read(./.env) means it can't read your secrets even if it tries. it's in git so the whole team gets it
.claude/settings.local.json
same idea but just for you, gitignored
.mcp.json
project mcp servers, so anyone who clones the repo gets the same tools
.claude/skills/name/SKILL.md
procedures you only need sometimes, like release steps or a migration checklist. it only loads when the task calls for it
.claude/agents/
subagents with their own prompt and tool list. a reviewer that can read but not edit is a good first one
hooks, in settings.json
anything that has to happen every single time. a PreToolUse hook that exits with code 2 blocks the call outright
how i'd split it: if claude ignoring it once is annoying, CLAUDE.md. if it'd be a disaster, hook. on a long task it kinda drifts from instructions, a hook just runs
when an agent run goes sideways it's usually the harness, not the model. what i'd have in the repo before letting claude code or codex run unattended:
1. AGENTS.md short enough to read every session. build, test and lint commands plus the mistakes it keeps making in this repo
2. one test command that exits nonzero on failure. if it can't run the check it can't know it's done
3. a definition of done written before the task starts and checked at the end, not vibes
4. a progress file it appends to after each step so a fresh session resumes without the old chat
5. a folder allowlist. touching anything outside it means stop and ask
6. a token or time budget per task so a stuck loop dies on its own
7. decisions and the why behind them somewhere every agent reads. mine live in tokenrip so claude, codex and cursor stop reopening the same argument
8. a different model reviews the diff, not the plan
most of this is an afternoon of setup tbh. save it for your next overnight run
9 of 43 runs, and in the main one it only went off script after the normal attempts failed. that's the moment i'd watch in my own agents, plan a dies and it starts improvising
i'd tie the stop rule to failures, not the clock. what stopped it was a human maintainer saying no
In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities.
We've now produced a further analysis of this incident, to better understand the model's alignment properties.
anthropic.com/research/align…
