I am rebuilding (unslopping) an app chaotically made by a non-dev friend using AI, that quickly got to 40ish paying users. This is an app to study for public job exams in a specific territory. Infamously topics are public but their content is not. He found a way to agregate from exams that are made public afterwards, and found sufficient overlapping to slowly build a db of questions and several types of tests: per topic, real exam simulacrum, an algo that tells you what to study, etc. From a dev pov the app is a mess though, includin g broken ui, ad hoc components pn top of libraries with both tailwind and hardcoded css, and database debauchery, etc. I am rebuilding the foundation, respecting his vision, let's see how many users fir the next exam :)
Several. Yesterday a friend with no prior coding experience or knowledge showed me an app he initially built to help him study for public administration job positions. The exams for this positions are public (spain), but the tools are scarce, expensive or he did not like. So he used lovable, then switched to web gemini and claude, then paid claude. He now has +130 very active users on an initial free tier, while he figures out. The app is on github, runs on vercel with supabase, react, tailwind, bun... he has no idea what he is doing. I even installed claude code for him, got him an ssh key so he can do it locally, etc.
Another: claude code cracked for me some software that was calling a home that did not exist anymore via headless ghidra.
Another: I am a teacher, and qualifications and feedback is very very time consuming, specially in loose workflows with several sources and tools that are not connected. During class presentations I take loose notes. Now I have a local folder where I drop my 1 student list, with names and emails, 2 my loose notes, and 3 a qualification & feedback sheet model; then claude creates a sheet per student, formats and copies the feedback to the right sheet cell, waits for my corrections, then sends everything to their school emails. Much easier, much less time consuming.
That said, I am very critical of AI, I align with voices and reports calling for AI companies to give back, as they took much, and for AI as public infrastructure, to an extent. I see datacenters as probably inevitably future public infrastructure, with a public model that could resemble that of electricity etc (in spain) or more public (less private). I am wary of the actual and future ecological and social impact of datacenter building and other problems AI is or will create. It is difficult to negate its usefulness, though: it is like having several very fast assistants with expert knowledge on several fields, that just get better every month. We will see.
Do your students know their feedback is AI generated?
A pre-AI example from the military : course reviews and performance assessments were changes to be from a set of about 70 descriptors instead of written custom. Instructors could modify them, but many didn't or did so only trivially. The system was junked in within three years because of the obvious: those giving feedback didn't own it, and those receiving feedback didn't value it.
Feedback is not AI generated, just structured grammar corrected and pasted at the right spot. I then review the output to see if it remains within what I meant to convey with my feedback. This is relatively simple feedback, I have 'highlight' and 'to improve' fieldsm, with a few sentences in each, very focused. If the feedback felt depersonalized or its meaning or tone changed I would fix it or not use this method. For longer forms of feedback it may not work.
> Now I have a local folder where I drop my 1 student list, with names and emails, 2 my loose notes, and 3 a qualification & feedback sheet model; then claude creates a sheet per student, formats and copies the feedback to the right sheet cell, waits for my corrections, then sends everything to their school emails
I've built a small system to do this anonymously. There is a students.csv with real data, a notes.txt that contains my unstructured comments and grades associated to ids (not names or student data), and a model.ods that contains the grading sheet model.
Claude takes the notes.txt and produces a json with corrected comments in the structure I asked for (highlights/needs work/grade), associated to student ids (not real data). This works both for single id or multiple id, for group assignments.
Then a script takes the json, creates a model sheet per student or group of students, fills the right cells, checks the ids against students.csv to fill the real names, and produces the pdf in a pdf/ folder.
Another script sends the pdfs.
I gitignore the sensible files, including an .env with SMTP pass, and denied permissions to Claude for those files using a rule at .claude/settings.json.
There is also a config file to change language, email text and other things.
I believe this is safe and compliant with GDPR, unless Claude ignores the deny rules! Any comments appreciated. Thanks.
possibly not or grey area under GDPR if I use identifiable information, as it is sent to Anthropic for processing, no matter if used for training or not, but I am unsure about this, I should probably anonymize and research it more, thanks for pointing it out
yes something like that, additionally most steps do not require data going through claude anymore, as it already wrote the script that take the student list and the qualifications model and produce a model per student, AND the script that takes that and sends each to its right email. The problematic part is when claude reads my notes and formats them into each of those student qualification sheets. There I would need some form of scrambling as you suggest, not to hijack the thread but ideas appreciated for a minimal setup. I believe claude respects .gitignore.
Maybe you could run a local script or smaller local model that takes a first pass through the notes and replaces every instance of a given name with their assigned number?
There is another institution I teach at that gives us Gemini, but not via API, which limits its use for this kind of work to an extent, I could do it via drive, I assume. There being a contract puts the institution and Google as responsible of the data. The first institution I was talking about has MS Teams, without AI afaik, but if they contract it I guess I can do the same with sharepoint, etc.
Sorry to tell you but it’s not grey area, it’s full on black. You do not have permission to share such data with a third party provider that doesn’t have strict privacy guarantees and that you have a data processing agreement with. TOS are not sufficient.
pairing this with rewind-like applications, it would seem that "user files" and particularly "user organized files" as a paradigm could also be completely unnecessary, as you recall states/activities/projects via intention and the agent presents you with the right abstraction to navigate those
There is a build in filesystem handling system. For example, if you ask it to build a Todo task agent it will need to persist the tasks and their state to disk. It knows which syscalls to use in the pneuma microkernal and it builds with that in mind. When you start the app it will reauesf access to the file system.
In the conversation I (mostly Claude, to be fair) was speculating about some caveats of fully on the fly generated ui/apps, in relation to ui/ux design and brand recognition, including how does ui/ux innovation enter the model if this is a paradigm shift, and what kind of job positions/skillsets would that paradigm destroy/create
If we fully embrace this generative software concept for the sake of this thread, then the UI/UX, is going to be optimized and personalized to your tastes, skills, affinities, and (dis)abilities. Brand comes into the equation as a proxy for trust, so if this whole scenario were to come true, maybe that's not so relevant anymore.
reply