Scattered on Purpose (@ScatteredOP) header
Scattered on Purpose (@ScatteredOP) profile picture

Scattered on Purpose

@ScatteredOP

Solo founder. 10+ products across unrelated domains. BFA grad turned AI-powered dev. The synoptic mind is the superpower. Weekly essays from the intersections.

 Everywhere, on purpose
scatteredonpurpose.substack.com
Joined March 2026
  • Tweets173
  • Following55
  • Followers4
Replying to @ScatteredOP
A local model can't make an idea original or create rights in it. It only changes who receives my unfinished work. It turned out to be four questions: what did I contribute, what rights exist, how do I make it distinctive, where does unfinished work go? scatteredonpurpose.substack.…
Image from Scattered on Purpose (@ScatteredOP)
2
2
7
"Just use open source" mixes up two different questions. For my purposes, the label answers nothing. I need three other answers: Where does it run? What does the app around the model send online? What does the license allow?
Checklist titled 'Three questions the label doesn't answer': Where does it run? What does the app around the model send online? What does the license allow?

ALT Checklist titled 'Three questions the label doesn't answer': Where does it run? What does the app around the model send online? What does the license allow?

2
Most AI privacy talk is about training, and the controls there are real. But turning training off is not the same as keeping a conversation on my own computer. The service still has to process everything I send it. scatteredonpurpose.substack.…
Banner: the line 'Training is only one switch.' over a faded paper-cut illustration of two wooden dye vats, one full of indigo and cream cloth, one nearly empty.

ALT Banner: the line 'Training is only one switch.' over a faded paper-cut illustration of two wooden dye vats, one full of indigo and cream cloth, one nearly empty.

3
Scattered on Purpose retweeted
Scattered on Purpose (@ScatteredOP) profile picture
Scattered on Purpose@ScatteredOPNov 9
A chat window is built for half-formed ideas. I describe a feature before I know if it belongs. I paste in a spec nobody else has seen. That's why the boring questions matter: where does that material go, who can process it, what happens after? Part 4 of a four-part series.
Image from Scattered on Purpose (@ScatteredOP)
1
2
2
10
Scattered on Purpose retweeted
Scattered on Purpose (@ScatteredOP) profile picture
Scattered on Purpose@ScatteredOPNov 9
A local model can't make an idea original or create rights in it. It only changes who receives my unfinished work. It turned out to be four questions: what did I contribute, what rights exist, how do I make it distinctive, where does unfinished work go? scatteredonpurpose.substack.…
Image from Scattered on Purpose (@ScatteredOP)
2
2
7
A chat window is built for half-formed ideas. I describe a feature before I know if it belongs. I paste in a spec nobody else has seen. That's why the boring questions matter: where does that material go, who can process it, what happens after? Part 4 of a four-part series.
Image from Scattered on Purpose (@ScatteredOP)
1
2
2
10
more replies
If I set one up, and this is a plan, not a report: start small, try it on something that doesn't matter, and check whether the interface quietly adds web search, cloud fallback, or sync. For SpiceShelf, unpublished specs would stay local. Public research could go to the cloud.
1
2
A decision record I plan to keep: the first concept note, what changed after talking to readers, and the model's suggestions I rejected. It shows whether the direction came from my choices or from accepting the most convincing answer over and over. scatteredonpurpose.substack.…
Banner: the line 'Keep the suggestions you rejected.' beside a paper-cut illustration of indigo test strips pinned to a cream wall.

ALT Banner: the line 'Keep the suggestions you rejected.' beside a paper-cut illustration of indigo test strips pinned to a cream wall.

4
Two shortcuts I don't trust. Asking a second model doesn't certify originality. It just widens the search. And asking a model to confirm its own suggestion has never existed isn't a test at all.
Paper-cut illustration of a person at a laptop showing a generic weather app, leaning in hopefully. Thought bubble: 'Has anyone built this before?' A bubble from the laptop answers: 'Nope! Totally original.'

ALT Paper-cut illustration of a person at a laptop showing a generic weather app, leaning in hopefully. Thought bubble: 'Has anyone built this before?' A bubble from the laptop answers: 'Nope! Totally original.'

1
1
8
In a 2024 study, short stories written with access to AI ideas were rated as more novel and more useful. They also came out more alike. So I want to ask two questions about any suggestion. Is this better? And is this more like everything else? scatteredonpurpose.substack.…
Bar chart of rating increases versus writers with no AI ideas: novelty up 5.4% with one AI idea and 8.1% with up to five; usefulness up 3.7% with one and 9.0% with up to five. Note: the stories also became more similar to each other.

ALT Bar chart of rating increases versus writers with no AI ideas: novelty up 5.4% with one AI idea and 8.1% with up to five; usefulness up 3.7% with one and 9.0% with up to five. Note: the stories also became more similar to each other.

1
1
10
The model knows a lot about readers in general. It doesn't know mine.
Handwritten quote on a taped index card: 'An imaginary reader's enthusiasm is not evidence that a real one wants the product.' Signed @ScatteredOP.

ALT Handwritten quote on a taped index card: 'An imaginary reader's enthusiasm is not evidence that a real one wants the product.' Signed @ScatteredOP.

1
1
7
Ask a model for ten variations, and you often get one idea in ten outfits. New names. New colors. And the user still does exactly the same thing. What I want instead: alternatives that actually work differently.
Two panels under the title 'Is it really a different idea?' The left panel, 'Looks different', shows a row of differently colored boxes: new names, new colors, and the user still does the same thing. The right panel, 'Works differently', shows a box and a circle: a different workflow and a different reason to open it.

ALT Two panels under the title 'Is it really a different idea?' The left panel, 'Looks different', shows a row of differently colored boxes: new names, new colors, and the user still does the same thing. The right panel, 'Works differently', shows a box and a circle: a different workflow and a different reason to open it.

1
1
12
Before I open a chat with a model, I want a short note written first. Who is this for? What do they struggle with? Which answers did I actually see, and which am I assuming? It gives me a fixed point to measure every suggestion against. scatteredonpurpose.substack.…
Banner: the line 'The note doesn't need to be right. It needs to exist.' beside a paper-cut illustration of hands binding plain cream cloth with string before it is dyed.

ALT Banner: the line 'The note doesn't need to be right. It needs to exist.' beside a paper-cut illustration of hands binding plain cream cloth with string before it is dyed.

1
1
19
When I type "help me design a book app" into a chat, I've handed over almost every decision before it starts. Audience, problem, features, feel. The model will do that well. That's the risk. I end up polishing a direction I never chose. Part 3 of a four-part series.
Image from Scattered on Purpose (@ScatteredOP)
1
1
2
13
more replies
Two shortcuts I don't trust. Asking a second model doesn't certify originality. It widens the search. Asking a model to confirm its own suggestion that has never existed isn't a test at all. What I want: a decision record, including the suggestions I rejected.
Image from Scattered on Purpose (@ScatteredOP)
1
2
5
None of this stops someone else from landing in the same place on their own. I could never prevent that. What the habits give me is better reasons. A model offers more possibilities than I can produce alone. I still need a reason to choose. scatteredonpurpose.substack.…
Image from Scattered on Purpose (@ScatteredOP)
1
2
13
Load more