See more of my writing here.
In this post, I'll start from scratch and build up to OpenClaw's architecture step by step, showing how you could have invented it yourself from first principles, using nothing but a messaging API, an LLM, and the desire to make AI actually useful outside the chat window.
End goal: understand how persistent AI assistants work, so you can build your own (or become an OpenClaw power user).
When you use ChatGPT or Claude in a browser, there are several limitations:
| #!/usr/bin/env bash | |
| # listgitfiles.sh | |
| # List raw GitHub URLs for files in a repo or subfolder, filtered by extensions. | |
| # Full README is embedded and printed only with: listgitfiles.sh -h|--help | |
| set -euo pipefail | |
| # ------------------------- Minimal Help (usage) ------------------------- | |
| print_min_help() { | |
| cat <<'USAGE' |
| // ==UserScript== | |
| // @name GoodCloud Dynamic Map | |
| // @namespace https://www.steveinnovates.com | |
| // @version 2025-07-24 | |
| // @description Displays a dynamic map of your bounded devices | |
| // @author Steve Oziel | |
| // @match https://www.goodcloud.xyz/ | |
| // @icon https://www.google.com/s2/favicons?sz=64&domain=goodcloud.xyz | |
| // @grant none | |
| // @downloadURL https://gist.github.com/gmoz22/3daa61753f27562dd7af460cb8a12eb6 |
This collection is limited to only include the reports that were submitted as security vulnerabilities to the curl bug-bounty program on Hackerone.
Several other issues not included here are highly suspcious as well.
tl;dr this demo shows how to call OpenAI's gpt-4o-mini model, provide it with URL of a screenshot of a document, and extract data that follows a schema you define. The results are pretty solid even with little effort in defining the data — and no effort doing data prep. OpenAI's API could be a cost-efficient tool for large scale data gathering projects involving public documents.
OpenAI announced Structured Outputs for its API, a feature that allows users to specify the fields and schema of extracted data, and guarantees that the JSON output will follow that specification.
For example, given a Congressional financial disclosure report, with assets defined in a table like this:
| function transform(clip) { | |
| const pathRegex = /^[a-z]+:\/\/[^:\/]+(:[0-9]+)?\/(.*?)(\/[0-9]+)?(\?.*)?$/ | |
| const uriPath = clip.text.replace(pathRegex, '$2'); | |
| const spotifyUri = "spotify:" + uriPath.replaceAll('/', ':') | |
| return spotifyUri; | |
| } |
| FROM hayd/alpine-deno:1.10.1 | |
| WORKDIR /src/app | |
| ADD deps.ts ./ | |
| RUN ["deno", "cache", "deps.ts"] | |
| ADD *.ts ./ | |
| RUN ["deno", "cache", "mod.ts"] | |
| ENTRYPOINT ["deno", "run", "--unstable", "--allow-net", "--allow-hrtime", "--allow-env", "--cached-only", "--no-check", "mod.ts"] |
| #!/bin/bash | |
| # Context: https://jamiehall.cc/2020/03/10/delete-all-your-tweets-with-one-line-of-bash/ | |
| # https://news.ycombinator.com/item?id=22689746 | |
| twurl "/1.1/statuses/user_timeline.json?screen_name=YOUR_TWITTER_HANDLE&count=200&max_id=$(twurl '/1.1/statuses/user_timeline.json?screen_name=YOUR_TWITTER_HANDLE&count=200&include_rts=1' | jq -r '.[9]|.id_str')&include_rts=1" | jq -r '.[]|.id_str' | parallel -j 10 -a - twurl -X POST /1.1/statuses/destroy/{1}.json > /dev/null |
Serverless infrastructure like AWS Lambda and Google Cloud Functions have made it much cheaper for developers to offer server-side code for public consumption without keeping a server always running.
If these functions could be declared as stateless or deterministic, costs can be brought down even more because only the first invocation needs to be executed. Cached response could be returned for future invocations with the same input arguments.
All modern browsers support URL lengths of thousands of characters, even on mobile. A lot of data can be embedded and passed around directly in the URLs (instead of passing identifiers which requires a look-up which costs server time).
So here's a thought:
