Prompt:
do this, adding the capbility to push the message via my discord as well as todoist. you can improve this any way you like but it should be lightweight, always on, robust to system restarts, not use too kuch claude token usage, and accurate (screen for false positives) - bascially i want to know the moment this goes live. # SA AI Royal Commission — Terms of Reference Watcher
## Objective
Build and install a watcher on this Mac Mini (M4, macOS) that polls public sources for the publication of the **terms of reference for South Australia's Royal Commission into Artificial Intelligence** (announced by Premier Malinauskas on 10 August 2026, ToR promised within 4-6 weeks, i.e. […]
What I was trying to achieve: I wanted to build an always-on monitoring system that alerts me the instant a major policy announcement goes live.
What happened: Built a regex-and-Sonnet verification pipeline that screens for false positives and sends Discord notifications, deployed to launchd. It combines cheap regex filtering with AI verification gates to balance cost and accuracy.
Reflection: This one worked out of the gate - got a test DM from Discord, and can already see it screening false positives.
Prompt:
ive read a lot about anthropics new shift to watermarking. AFAIK it is mostly to comply with EU regulations. i am seeing it touted as a move in the direction of detection and i disagree for several reasons: 1) it is not about building tools to detect AI content it is about watermarking output 2) it is a technique that os easily obfuscated 3) it seems designed to address the compliance conerns more than anything else and it would catch bot spam but not a himan user who had done any obfuscation methods and so on. Already the socials are filled with hot takes and misinterpretations. in the writing projects folder i would like you to make a nice looking single file explainer of how watermarking […]
What I was trying to achieve: I wanted to build an interactive explainer that challenges misconceptions about AI watermarking and clarifies why it's compliance-focused rather than detection-focused.
What happened: Built an interactive tool that breaks down the technical constraints and social misinterpretations of watermarking, deployed to Tailscale for mobile review. The explainer walks through why it won't catch obfuscated content and who it actually serves.
Reflection: Not quite finished - the v1 came out a little too technical
Prompt:
i just dropped a voice memo into the folder sync/tools/thought-transcriber - this voice memo is the first example of what this tool should be, and contains instructions for building itself as well as some examples of what the tool will actually do. Transcribe and QA the transcript, build it, run it.
What I was trying to achieve: I wanted to build a tool that automatically processes voice memos into cleaned transcripts and structured action items.
What happened: Built a voice-memo watcher with transcription QA, content extraction, and outputs to Buffer/Todoist/Discord, persisted with launchd. Verified it survives system restarts and sends reliable Discord notifications.
Reflection: This one was a weird experiment. Can the AI transcribe a voice memo that contains instructions for building a tool that transcribes voice memos... Inception software.
Prompt:
is it possible to hitch you up to google search console?
What I was trying to achieve: I wanted to connect Google Search Console to Claude so I get notified about interesting search metric changes.
What happened: Built a zero-LLM-token monitoring agent using statistical anomaly detection to catch spikes, new pages, and rank movements, then sends Discord alerts. Runs on launchd with no ongoing token cost.
Reflection: Yes, it's possible, and ~20 minutes later I have the best data on website visits I've had in 16 years of running the site
Prompt:
i just plugged a data cable from the solar panels into my starlink router justbwindering if you can see it?
What I was trying to achieve: I wanted to know whether Claude could see the solar system I'd just plugged into the home network — and if it could, do something useful with it.
What happened: Claude found the Victron GX device on the network, and by the end of the session there was an always-on solar dashboard on the private tailnet — financials, a virtual fuel gauge, queryable history — hardened to survive reboots and locked away from accidental network exposure.
Reflection: "justbwindering" is a particularly good typo. This worked and is much better than the default platform.
Prompt:
i already have the chat dashboard which keeps a running memory thread across macbook and mac mini and uses Willisons transcription repo to prettify terminal code chats. im thinking a useful thing might be a "My Week with AI" semi-aitomated diary that gets published almost like an rss feed on my website which has highlights of the ways ive used AI, prompts, use cases with explanations and snippets of cose or whatever. im thinking it looks like: date, time, title, main prompt/s, what i was trying to achjeve, platform (claude code or claude app) model/s used, reflection. it would need to be selective […]
What I was trying to achieve: I wanted a semi-automated public diary of how I actually use AI — harvested from my chat logs each week, filtered for anything sensitive, reviewed by me on a Friday, and published plainly on my site with RSS.
What happened: Claude bolted a small pipeline onto my existing chat dashboard: a hard denylist plus a Haiku triage pass screens the week's sessions, then a Friday launchd job drafts the entries — prompts quoted verbatim, reflections left blank for me — and DMs me to review. Nothing publishes without my sign-off. The first draft included the session that built it.
Reflection: it's this...