I built Debloat: a free open-source Mac app that shows what's eating your memory and disk, and lets you fix it.
• Stop leftovers from closed apps
• See and stop Claude Code background tasks
• Move big files to an SSD in one click
• Block startup junk
https://t.co/2t6225TkdB
@wangyan_im Good call, those are different checkpoints. Today it only verifies the file count before deleting, which isn't enough. I'm adding per-file checksum verification, with each stage shown in the UI, so nothing local gets removed until the destination provably matches. Shipping soon.
Prompt caching tip nobody tells you: order matters.
The cache only matches from the start of the prompt. Static stuff first (system prompt, tools, examples), changing stuff last (user message, timestamps, IDs).
One timestamp at the top = zero cache hits, full price every call.
@apratxm lmao fair, if your prefill changes every call the cache can't save you. the trick is making the first 90% of the prompt boringly identical every time
@Modelflaredev 98.3% of input tokens came from cache. it's a coding agent, so every turn re-sends the whole convo and the prefix barely changes. funny part: cache writes were only 1.7% of tokens but 34% of the spend
I logged 7,407 of my own LLM API calls and priced them at list rates.
Without prompt caching: $6,301
With caching: $566
Same calls, same outputs. That's 91% of the bill gone. 🧵
@buildwithsid that's sick. full history or some kind of summarised memory? feels like the hard part is someone who's been texting it for 6 months, both for recall and for the token bill lol
@GergelyOrosz "It's not DNS" finally happened and it still took 20 hours. Moving a legacy system mid-outage as the fix is wild though. That's changing the engine while the plane is on fire.
@GergelyOrosz The AI equivalent of keeping a backup charger in every bag. Same goes for anyone building on these APIs: if your product calls one provider with no fallback, their outage is your outage.
@rauchg durable-harness being reusable is the big one. Most agent frameworks treat a crash mid-run as "start over and pay for every LLM call again". Resuming from the last completed step saves both time and money.
@TurnerNovak Matches my experience with coding agents. A cheaper model that needs three retries isn't cheaper once you count your own time fixing its output. Per-token price is the wrong metric, cost per finished task is the right one.
Full breakdown: https://t.co/mx6atYcCeC
Rough estimate of your own bill: https://t.co/dfqe5oaDDM
If your API bill is growing faster than revenue, DM me. I'll take a free look.
Other places I usually see money leak:
• easy requests going to the expensive model
• full chat history re-sent every turn
• the same question answered over and over with no cache