Solty

18 / one claude code trick a day / agents that work while i sleep / dm open

[Solty @0xSolty]

https://t.co/fixZv3pBdk

[Solty @0xSolty]

the prompt I used. paste it into Claude with your App_Privacy_Report.ndjson:

This is my own iPhone's App Privacy Report. Read only.

1. Write a small Python script that replays the whole log as a fast live feed in the terminal: one line per event with time, app, what it accessed (location, photos, camera, mic, contacts) or which domain it contacted. Color by type, mark known ad and tracker domains in red.

2. When it ends, print a summary: which apps checked my location and how often, which touched photos, mic or camera, which talked to the most tracker domains.

3. Tell me exactly which permissions to take back and where in Settings.

runs on iPhone in a-Shell (free): python3 https://t.co/31nUqbcw38

[Solty @0xSolty]

Claude Code read my iPhone’s hidden privacy log.

9 apps were watching me. one of them I forgot I even had

it started with the battery. dead by 3pm every day. warm in my pocket with the screen off

I thought it was just old. it wasn’t

there's a setting buried at the very bottom of Privacy & Security. App Privacy Report

it quietly logs every time an app touches your:

→ location
→ photos
→ mic
→ camera
→ and every server it talks to
almost nobody turns it on

12:03, I turned it on. put the phone in my pocket. lived a normal day

16:00, I opened it

the list didn’t fit on the screen

so I exported it, opened my laptop and gave Claude Code one prompt: replay my day and tell me who was watching

it started printing. it didn’t stop for 2 minutes

→ 2,340 events in 4 hours
→ 9 apps checked where I was
→ 6 apps opened my photos
→ 47 different tracker servers
→ 3 apps were busiest while I wasn't even touching the phone

then Claude printed the summary. one name was at the top of every list

an old puzzle game

I haven’t opened it since March. it’s not even on my home screen anymore. I forgot it existed

today it asked for my location 48 times. every 5 minutes. on the bus, in the shop, at home

it talked to 23 servers. 17 of them were ad trackers I’d never heard of

at 14:37 it woke up on its own. phone in my pocket. screen off

it knew where I was all day. I didn’t even know it was still there

I deleted it. took back 14 permissions from 8 other apps. 5 minutes

by 10pm my battery was at 40%. it used to be dead by 3

your check tonight takes 3 minutes:

App Privacy Report → Settings → Privacy & Security → turn it on, look tomorrow
location → nothing needs “Always” except maps and Find My
photos → only your camera and gallery need “Full Access”
mic and camera → if you can’t say why an app needs it, turn it off
apps you forgot → if you haven’t opened it this month, it shouldn’t be watching you

found something? screenshot the report first, then take the permission back

save this. check your phone tonight. then check your mom’s

the Opus 5.5 prompt I used ↓

[Solty @0xSolty]

https://t.co/fixZv3pBdk

[Solty @0xSolty]

@eng_khairallah1 anytime man, let opus plan and sonnet type, your limit will thank you

[Solty @0xSolty]

Opus 5.5 is the best planner you'll ever pay for, and the most expensive typist. one model setting splits the job:

/model opusplan

plan mode → Opus 5.5 thinks
execution → Sonnet 5.5 writes the code
it switches by itself the moment you leave plan mode

the whole loop:
→ Shift+Tab into plan mode
→ Opus 5.5 reads the repo, picks the approach, writes the plan
→ you approve the plan
→ Sonnet 5.5 builds it in the same session, same context

the math on one long session (API prices, no cache):
all on Opus 5.5: ~$5.60
opusplan: ~$3.08

Opus 5.5 costs $4 / $20 per million tokens, Sonnet 5.5 costs $2 / $10. about 90% of a session is execution, so most of your tokens drop to half price. on Pro/Max that means your limit lasts longer

both phases get the 1M context window on the API. if yours doesn't, use /model opusplan[1m]

the full tree
> /model opusplan
> plan mode: Opus 5.5 plans
> you approve
> execution: Sonnet 5.5 builds
> stuck? Shift+Tab back into plan mode and Opus takes over again

paste this into Claude Code ↓

"Switch me to /model opusplan. For this task: enter plan mode, read only what you need, and write a plan with the files to touch, the risks and a test checklist. Stop and wait for my approval. After I approve, implement it step by step and run the tests after each step."

↳ https://t.co/ZZ2a6tbHED

[Solty @0xSolty]

https://t.co/fixZv3pBdk

[Solty @0xSolty]

@RohOnChain appreciate you bro, opus plans and sonnet types, best combo rn

[Solty @0xSolty]

Opus 5.5 is the best planner you'll ever pay for, and the most expensive typist. one model setting splits the job:

/model opusplan

plan mode → Opus 5.5 thinks
execution → Sonnet 5.5 writes the code
it switches by itself the moment you leave plan mode

the whole loop:
→ Shift+Tab into plan mode
→ Opus 5.5 reads the repo, picks the approach, writes the plan
→ you approve the plan
→ Sonnet 5.5 builds it in the same session, same context

the math on one long session (API prices, no cache):
all on Opus 5.5: ~$5.60
opusplan: ~$3.08

Opus 5.5 costs $4 / $20 per million tokens, Sonnet 5.5 costs $2 / $10. about 90% of a session is execution, so most of your tokens drop to half price. on Pro/Max that means your limit lasts longer

both phases get the 1M context window on the API. if yours doesn't, use /model opusplan[1m]

the full tree
> /model opusplan
> plan mode: Opus 5.5 plans
> you approve
> execution: Sonnet 5.5 builds
> stuck? Shift+Tab back into plan mode and Opus takes over again

paste this into Claude Code ↓

"Switch me to /model opusplan. For this task: enter plan mode, read only what you need, and write a plan with the files to touch, the risks and a test checklist. Stop and wait for my approval. After I approve, implement it step by step and run the tests after each step."

↳ https://t.co/ZZ2a6tbHED

[Solty @0xSolty]

@mirku21 is doctor promt audit live now or still rolling out, cant find it in my cli yet?

[Solty @0xSolty]

my friend's old Android sat in a drawer for 2 years. now it runs his whole flat

no new gadgets, one prompt. the full setup and the exact prompt are below ↓

5 minutes before he gets home he taps one button in Telegram:

/home

and by the time he opens the door:
> the lights are on
> the AC has already cooled the room
> his playlist is playing
> his PC is booted and on the desktop

the phone that did all of it never left the shelf. it's plugged in, on Wi-Fi, running one small script Opus 5.5 wrote for him

how it works:
→ Termux turns the old phone into a tiny Linux box
→ a Telegram bot sits there and waits for /home
→ Wake-on-LAN sends one packet and the PC wakes up
→ smart plugs and the AC get their commands over local Wi-Fi
→ the music starts on the speaker in the room
→ no cloud, no subscription, no new gadgets except a couple of cheap plugs

paste this into Claude Code ↓

"I have an old Android phone with Termux on my home Wi-Fi. Write a Python Telegram bot that runs there and on /home:
1. sends a Wake-on-LAN packet to my PC
2. turns on my lights and AC
3. starts my music playlist on my speaker
4. replies with what turned on
First ask me which lights, AC and speaker I have, then pick the right way to control each one. Only respond to my Telegram user ID. Add /off and make the bot start when the phone reboots. Walk me through the Termux setup one step at a time."

best upgrade a drawer phone ever got

[Solty @0xSolty]

my friend's old Android sat in a drawer for 2 years. now it runs his whole flat

no new gadgets, one prompt. the full setup and the exact prompt are below ↓

5 minutes before he gets home he taps one button in Telegram:

/home

and by the time he opens the door:
> the lights are on
> the AC has already cooled the room
> his playlist is playing
> his PC is booted and on the desktop

the phone that did all of it never left the shelf. it's plugged in, on Wi-Fi, running one small script Opus 5.5 wrote for him

how it works:
→ Termux turns the old phone into a tiny Linux box
→ a Telegram bot sits there and waits for /home
→ Wake-on-LAN sends one packet and the PC wakes up
→ smart plugs and the AC get their commands over local Wi-Fi
→ the music starts on the speaker in the room
→ no cloud, no subscription, no new gadgets except a couple of cheap plugs

paste this into Claude Code ↓

"I have an old Android phone with Termux on my home Wi-Fi. Write a Python Telegram bot that runs there and on /home:
1. sends a Wake-on-LAN packet to my PC
2. turns on my lights and AC
3. starts my music playlist on my speaker
4. replies with what turned on
First ask me which lights, AC and speaker I have, then pick the right way to control each one. Only respond to my Telegram user ID. Add /off and make the bot start when the phone reboots. Walk me through the Termux setup one step at a time."

best upgrade a drawer phone ever got

[Solty @0xSolty]

https://t.co/fixZv3pBdk

[Solty @0xSolty]

your 5-hour limit doesn't die while you work. it dies when you come back from a break and type <continue> one variable fixes most of it:

export CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS=100000

the prompt cache lives about an hour. leave a long session longer than that, and your next message re-reads the whole thing at full price.

since this month Claude Code fights back:

→ at ~54 min idle it compacts the session, right before the cache expires
→ by default only for sessions over 200K tokens
→ the variable drops that to 100K, so mid-size sessions get saved too

check it's on: /config → Idle compaction

the variable is one piece. the full setup has 3 more:

→ the resume warning, shipped yesterday. resume an old chat and Claude Code tells you when it will eat a big part of your 5-hour limit. pick "start new" and just mention the old chat. Claude looks it up instead of reloading all of it
→ CLAUDE_CODE_AUTO_COMPACT_WINDOW=200000. write the full number. "200k" was silently turning into a 100K window until this week's fix
→ /usage at the end of a long day. it now flags max effort, workflows, always-on context and huge tool results, the usual limit killers

the full tree
> long session goes idle
> ~54 min: idle compaction, before the cache expires
> sessions from 100K included with the variable
> resume warning before an expensive reload
> /usage shows what actually ate the limit

paste this into Claude Code ↓

"Audit my Claude Code setup for usage-limit leaks:
1. Add CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS=100000 to my shell profile.
2. Check /config and make sure Idle compaction is on.
3. If CLAUDE_CODE_AUTO_COMPACT_WINDOW is set anywhere, rewrite it as a full number (200000, not 200k).
4. Add one line to CLAUDE.md: when I come back to an old task after a break, start a new session and reference the old one instead of resuming it.
Show every change as a diff first. Don't apply anything until I say go."

↳ https://t.co/o5JUshaeFG

the model isn't what's eating your limit. the coffee break is.

[Solty @0xSolty]

your 5-hour limit doesn't die while you work. it dies when you come back from a break and type <continue> one variable fixes most of it:

export CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS=100000

the prompt cache lives about an hour. leave a long session longer than that, and your next message re-reads the whole thing at full price.

since this month Claude Code fights back:

→ at ~54 min idle it compacts the session, right before the cache expires
→ by default only for sessions over 200K tokens
→ the variable drops that to 100K, so mid-size sessions get saved too

check it's on: /config → Idle compaction

the variable is one piece. the full setup has 3 more:

→ the resume warning, shipped yesterday. resume an old chat and Claude Code tells you when it will eat a big part of your 5-hour limit. pick "start new" and just mention the old chat. Claude looks it up instead of reloading all of it
→ CLAUDE_CODE_AUTO_COMPACT_WINDOW=200000. write the full number. "200k" was silently turning into a 100K window until this week's fix
→ /usage at the end of a long day. it now flags max effort, workflows, always-on context and huge tool results, the usual limit killers

the full tree
> long session goes idle
> ~54 min: idle compaction, before the cache expires
> sessions from 100K included with the variable
> resume warning before an expensive reload
> /usage shows what actually ate the limit

paste this into Claude Code ↓

"Audit my Claude Code setup for usage-limit leaks:
1. Add CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS=100000 to my shell profile.
2. Check /config and make sure Idle compaction is on.
3. If CLAUDE_CODE_AUTO_COMPACT_WINDOW is set anywhere, rewrite it as a full number (200000, not 200k).
4. Add one line to CLAUDE.md: when I come back to an old task after a break, start a new session and reference the old one instead of resuming it.
Show every change as a diff first. Don't apply anything until I say go."

↳ https://t.co/o5JUshaeFG

the model isn't what's eating your limit. the coffee break is.

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Solty @0xSolty]

It is becoming f*cking dangerous

A solo founder builds his whole product alone with 7 Claude Code agents, and none of them is allowed to say "done" without proof

He opens the team board:
- A product owner that sets priorities and hands out the work
- One developer, the only agent allowed to touch the code
- A UX reviewer that tests the live build instead of trusting the developer
- A marketer that drafts everything but can't publish a thing
- A customer success agent that reads feedback and can't change anything
- All of it on one 4-core server and a flat Claude subscription

The trick is the rule: no status reports, only proof. A screenshot, a test run, a working link, or it didn't happen

A startup would hire 7 people for this. He runs it from one server

Same agent-team setup that keeps blowing up this week. Bookmark the full breakdown

[Solty @0xSolty]

It is becoming f*cking dangerous

A solo founder builds his whole product alone with 7 Claude Code agents, and none of them is allowed to say "done" without proof

He opens the team board:
- A product owner that sets priorities and hands out the work
- One developer, the only agent allowed to touch the code
- A UX reviewer that tests the live build instead of trusting the developer
- A marketer that drafts everything but can't publish a thing
- A customer success agent that reads feedback and can't change anything
- All of it on one 4-core server and a flat Claude subscription

The trick is the rule: no status reports, only proof. A screenshot, a test run, a working link, or it didn't happen

A startup would hire 7 people for this. He runs it from one server

Same agent-team setup that keeps blowing up this week. Bookmark the full breakdown

[Solty @0xSolty]

https://t.co/fixZv3pBdk

[Solty @0xSolty]

@gippp69 might actually ship something dumb just for the chipotle tier ngl

[Solty @0xSolty]

@RohOnChain saving this before i lose it in my bookmarks graveyard

[Solty @0xSolty]

126K people starred graphify. almost nobody turned on the part that changes how Claude Code works:

graphify claude install

without it, every task starts the same way:

→ Glob the whole repo
→ Grep for "auth"
→ open 30 files to find the 3 that matter
→ half the context gone before the first edit

with it, a PreToolUse hook fires before every Glob and Grep and tells Claude to read GRAPH_REPORT.md first. one page: the god nodes, the communities, and the weird connections nobody ever wrote down.

so Claude opens the map, then the 3 files. not the 30.

setup, 4 lines:
→ pip install graphifyy (two y's, the plain name is taken)
→ graphify install
→ /graphify . inside Claude Code. you get graph.html, GRAPH_REPORT.md and graph.json
→ graphify claude install. this writes the hook into settings.json

the repo claims up to 71.5x fewer tokens per query. test it on your own repo before you believe a number.

the full tree
> /graphify . turns the repo into a graph
> GRAPH_REPORT.md: one page about the whole codebase
> PreToolUse hook: read the report before any Glob or Grep
> Claude opens only the files the graph points to
> re-run /graphify . after big changes

paste this into Claude Code ↓

"Set up graphify in this repo:
1. Run pip install graphifyy, then graphify install.
2. Run /graphify . and show me the top 5 god nodes from GRAPH_REPORT.md.
3. Run graphify claude install and show me the hook it adds to settings.json before saving it.
4. Add one line to CLAUDE.md: read GRAPH_REPORT.md before searching the codebase, and re-run /graphify . after any change that touches more than 10 files.
Show every change as a diff first. Don't apply anything until I say go."

↳ https://t.co/kV27bgg61H

[Solty @0xSolty]

126K people starred graphify. almost nobody turned on the part that changes how Claude Code works:

graphify claude install

without it, every task starts the same way:

→ Glob the whole repo
→ Grep for "auth"
→ open 30 files to find the 3 that matter
→ half the context gone before the first edit

with it, a PreToolUse hook fires before every Glob and Grep and tells Claude to read GRAPH_REPORT.md first. one page: the god nodes, the communities, and the weird connections nobody ever wrote down.

so Claude opens the map, then the 3 files. not the 30.

setup, 4 lines:
→ pip install graphifyy (two y's, the plain name is taken)
→ graphify install
→ /graphify . inside Claude Code. you get graph.html, GRAPH_REPORT.md and graph.json
→ graphify claude install. this writes the hook into settings.json

the repo claims up to 71.5x fewer tokens per query. test it on your own repo before you believe a number.

the full tree
> /graphify . turns the repo into a graph
> GRAPH_REPORT.md: one page about the whole codebase
> PreToolUse hook: read the report before any Glob or Grep
> Claude opens only the files the graph points to
> re-run /graphify . after big changes

paste this into Claude Code ↓

"Set up graphify in this repo:
1. Run pip install graphifyy, then graphify install.
2. Run /graphify . and show me the top 5 god nodes from GRAPH_REPORT.md.
3. Run graphify claude install and show me the hook it adds to settings.json before saving it.
4. Add one line to CLAUDE.md: read GRAPH_REPORT.md before searching the codebase, and re-run /graphify . after any change that touches more than 10 files.
Show every change as a diff first. Don't apply anything until I say go."

↳ https://t.co/kV27bgg61H

[Solty @0xSolty]

https://t.co/fixZv3pBdk

[Solty @0xSolty]

@RohOnChain it useful, bookmark

[Solty @0xSolty]

@RohOnChain thanks Roan

[Solty @0xSolty]

I asked Claude to check the new pump fun promo before I put a dollar in

the deal:
→ sign up + KYC
https://t.co/bodXbtD55o
→ deposit $50+ (card or crypto card)
→ make some volume every day for 6 days, even $2-3 is enough
→ $8.35 in rewards for each day

the math: 6 × $8.35 = $50.10 in rewards. the same size as your deposit, and the deposit stays yours.

the trick: buy any coin and sell it right away. you only lose a few cents in fees per day. the one rule is don't skip a day.

[Solty @0xSolty]

https://t.co/fixZv3pBdk

[Solty @0xSolty]

/loop and /remote-control both freeze when you close the lid. this one keeps running:

claude --cloud "fix the failing tests"

the difference is where the session lives:

→ /loop runs on your laptop. lid closed, loop frozen
→ /remote-control turns your phone into a remote. the session still sleeps with your laptop
→ claude --cloud spins up a VM on Anthropic's side, clones your repo and keeps going

it's hidden from claude --help, that's why most people never found it.

in the morning:
→ /teleport (or /tp) pulls the session into your terminal with the branch and the full history
→ /tasks then t does the same from the task list

want it every night? /schedule → "Cloud schedule". each run is its own isolated session in the cloud.

and this week cloud sessions got faster to start: ~6s instead of ~17s, because they warm up while you're still typing.

before you close the lid:
→ push your commits. the VM clones from GitHub, not from your disk
→ log in with /login. API keys don't work here, Pro/Max/Team/Enterprise do
→ /teleport wants a clean working tree and the same repo

the full tree
> your laptop: write the task, push, close the lid
> claude --cloud: VM clones the repo and runs it
> /schedule: the same thing every night on a timer
> /teleport: branch + history back in your terminal
> /remote-control: only for sessions that live on your machine

paste this into Claude Code before you go to bed ↓

"I'm about to close my laptop. Get this task ready for a cloud session:
1. Commit and push everything on this branch.
2. Write TASK.md with the goal, the files involved, and how to check it's done (which tests must pass).
3. Add one rule to CLAUDE.md: cloud sessions run the full test suite before finishing and write a short summary to HANDOFF.md.
4. Give me the exact claude --cloud command to start it.
Show every change first. Don't push until I say go."

↳ https://t.co/HyaGXzN8HA

your laptop needs sleep. your agent doesn't.