Building with AI. Sharing the useful bits.
Seoul
Joined January 2022
- Tweets218
- Following21
- Followers12
- Likes42
I asked Claude Code why my Mac only had 26GB free.
The biggest surprise wasn't my code. It was Chrome.
17GB in Chrome's folder. 4GB of that is a single file: weights.bin, Chrome's on-device AI model. I never knowingly installed it.
What else it found:
• Xcode DerivedData: 6.3GB
• ~/.cache: 6.4GB
• npm cache: 4.2GB
~21GB of rebuildable stuff, found in about 2 minutes.
Check yours: chrome://on-device-internals
If you want that space back:
• Chrome: Settings > System > turn off "On-device AI". Deleting weights.bin alone doesn't stick, Chrome downloads it again
• Note: some Chrome features that rely on the model (incl. some scam detection) stop working
Chrome's own docs: the model needs 22GB free to download and is removed if free space drops below 10GB
developer.chrome.com/docs/ai…
The most useful part of this recap is about your usage limit.
From the Claude Code changelog:
• Idle compaction: a 200K+ token conversation left idle gets summarized. If the prompt cache (~1h) has expired, your next message uses the summary instead of reprocessing everything
• On Pro and Max, resuming an inactive conversation now warns you when it would use a noticeable share of your 5-hour limit
• You can switch idle compaction off or on in /config
How I'd use it:
1. Back after a long break? Take the warning seriously. A new chat that refers to the old one is often cheaper
2. One hard problem? Type ultrathink in that message. It runs just that turn at xhigh effort, no need to raise effort for the whole session
Sources:
Claude Code changelog (2.1.302, 2.1.304, 2.1.311, 2.1.314): code.claude.com/docs/en/chan…
Computer use feels clumsy in Claude Code if you point it at websites. That's the wrong tool for that job.
From the computer-use docs: under screen control, browsers are view-only. Claude can look, not click.
For the web, use Claude in Chrome instead:
1. Install the Claude in Chrome extension
2. Start Claude Code with: claude --chrome
3. Run /chrome → "Enabled by default" so you never need the flag
It uses your real Chrome logins, can attach files up to 10 MB, and pauses for you at logins and CAPTCHAs.
Save computer use for what Chrome can't reach: native Mac apps, the iOS Simulator, GUI-only tools. Turn it on in /mcp → computer-use (macOS, Pro/Max/Ultra).
What tripped me up:
• Browser tools dead after a long idle? The extension's service worker slept. /chrome → Reconnect extension
• Every command hangs? A JS alert or confirm dialog (for me, a "Leave site?" prompt) is blocking the page. Dismiss it by hand
• Upload refused? Over 10 MB, or a file with multiple hard links (common in node_modules). Copy it first
One line in Claude Code's best-practices page changed how I run sessions:
"Every request sends the whole conversation, so a session that's collected three unrelated tasks re-reads all three on every turn of the fourth."
What I do now:
• /clear between unrelated tasks. It costs nothing, and the next task stops paying for the last one's history
• /clear <name> names the session I'm leaving, so I can come back to it
• On a long session, /compact with a focus ("keep the API changes and the test commands"). Plain /compact is itself a big request
• On paid plans, /usage flags long context or frequent cache misses once either passes 10% of your usage
What's your rule for when to /clear vs keep going?
The number in OpenAI's announcement that matters more than the $172K bounty: 25%.
From their announcement page:
• Their last security sprint put 25% of product engineers on finding and fixing weaknesses
• Using Codex, engineers investigated findings, reproduced vulnerabilities in isolated environments, and prepared patches for review
• $1B in subsidized access over six months for frontline defenders protecting essential services
The part small teams can copy today is the loop: find → reproduce in isolation → patch → a human reviews.
The model does the grind. A person still approves the fix.
As AI becomes more capable at finding vulnerabilities, defenders need to put it to work before attackers do.
We’re planning our next internal security sprint at OpenAI and partnering with Cathedral Systems, whose researchers will use our models alongside our teams to challenge our defenses. They’ll bring differentiated expertise and approaches to our ongoing work to find and fix weaknesses with AI.
We’re also raising bug bounty rewards, including more than quadrupling rewards for critical findings. Exceptional findings will be eligible for up to $172,000, including bonuses.
We’ll keep advancing our models, using them to strengthen our own systems, and expanding access for trusted defenders.
openai.com/index/preparing-o…
Source: OpenAI, "Preparing our next security sprint to harden OpenAI" (Nov 2, 2026)
openai.com/index/preparing-o…
One-line prompts like this work. The hidden cost is stacking them in one chat.
From the Claude Code best-practices page:
"Every request sends the whole conversation, so a session that's collected three unrelated tasks re-reads all three on every turn of the fourth."
It's discounted while the cache is warm, full price once it expires.
What I do now:
• /clear between unrelated tasks. It costs nothing, and the next task stops paying for the last one's history
• /clear extension-build names the session I'm leaving, so I can come back to it
• Run /usage. On paid plans it flags long context or frequent cache misses once either passes 10% of your usage
Source: Claude Code docs, "Best practices" → "Keep cost and context down"
code.claude.com/docs/en/best…
The detail this post leaves out: $350K–$500K a year.
From Anthropic's announcement page:
• Placements of 6 weeks, 3 months, or 6 months
• SF preferred, remote possible for shorter ones
• 1:1 mentor + a dedicated compute budget
• No PhD, ML experience, or papers required
But "no ML background" doesn't mean anyone. They want people who've already led ambitious technical work: quantitative professors, technical founders, staff-level engineers.
Their reasoning is the interesting part: what matters most in AI research now is framing the agenda and judging what's worth doing. That's the skill they're paying for.
Deadline: Nov 22.
Applications are now open for the Anthropic Residency, a paid, fixed-term research program for academics, technical founders, and staff-level engineers who want to make AI safe and beneficial for everyone. No ML background is required.
Apply by November 22: alignment.anthropic.com/2026…
Source: Anthropic's Alignment Science blog, "Anthropic Residency: applications open for established technical leaders"
alignment.anthropic.com/2026…
Turning on /fast halfway through a long Claude Code chat is the most expensive way to use it.
From Anthropic's fast mode docs:
• It bills your usage credits directly, even if your plan still has usage left
• The first time you turn it on in a chat, the whole conversation is charged at the uncached fast-mode price
• Fast and standard keep separate caches, so each flip re-reads part or all of the chat uncached
• Once on, it stays on in new sessions by default
My rules now:
- Decide at the start of a session, never mid-way
- Check for the ↯ next to the prompt before a long run
- Still using fast on Opus 4.8? It's removed Nov 30 and requests will error, not fall back. Opus 5.5 fast is cheaper anyway ($8/$40 vs $10/$50 per MTok)
Sources: Anthropic's Claude Code docs, "Understand the cost tradeoff"
code.claude.com/docs/en/fast…
Opus 4.8 fast mode removal (Nov 30, no fallback) is in the Claude Code changelog, v2.1.316
You stepped away from Claude Code. Your usage limit didn't.
Anthropic's docs list what still sends your full context while a session sits idle:
• A /loop or scheduled task fires on its interval, even with nobody there
• A message from another of your sessions arrives as a new turn
• An active /goal checks in on background work, up to 3 times between your prompts
Before I walk away now:
- Stop any loop I'm not actually watching
- Want a fresh start later? /clear costs nothing. /compact on a big chat is itself a big request
- Run /usage. If the footer says "f to fix", press f and it applies the fixes it can
Source: Anthropic's Claude Code docs, "Why usage climbs in a long session"
code.claude.com/docs/en/cost…
The "f to fix" shortcut needs v2.1.305+. The 3-check-in cap on /goal is v2.1.246+ (uncapped before). Check yours with claude --version
Claude Code has a flag that isn't in --help.
claude --advisor opus
Your cheaper main model does the work. A stronger model reviews the whole conversation at the moments that decide the outcome. It tends to step in:
→ before committing to an approach
→ when the same error keeps coming back
→ before declaring the task done
Small things the docs mention:
- Haiku main + Opus advisor is the lowest-cost pairing with strong planning
- Type "consult the advisor before you continue" to call it on demand
- Toggling /advisor mid-session keeps your prompt cache
- On a subscription, advisor tokens still count toward your limits
Experimental, Anthropic API only.
Heads-up: resuming an old Claude Code chat can quietly eat a big chunk of your 5-hour limit.
Here's why, from Anthropic's own docs:
• On a Pro/Max plan, your main chat's cache expires after 1 hour idle
• Resume after that and the whole conversation is re-processed with no cache hits
• Running /compact right after resuming is the most expensive move
Claude Code now warns you before this happens (v2.1.304+).
What I do now:
- Idle 1h+? Start a new chat. Claude can look up the old one if you mention it
- Mid-session, /rewind instead of /compact to drop a dead end. It lands on a prefix that's still cached
- Check /usage → "Prompt cache (main)" shows your hit ratio
Sources:
Prompt caching docs: code.claude.com/docs/en/prom…
Changelog (2.1.304 resume warning): code.claude.com/docs/en/chan…
