Rendered at 20:47:19 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
muzani 1 days ago [-]
Movies screenwriters seem to have found the perfect balance.
A character description looks like this: "APOLLO CREED. Creed is twenty-eight years old. He is a tall, smooth-muscled Black with barely a scar on his light coffee-colored face."
It includes everything important for casting this guy. It doesn't say what he's wearing, hair style, things like a wedding ring. Once they cast an actor, the actor fills this up.
Fight scenes are designed to be as fast to read as the action. Poor writing is something like, "Terrorists B and C fire RPGs at the van. The van makes evasive maneuvers. After the third rocket, the van flips off road." It doesn't make the story nor the scene clearer.
The better script: "The van takes evasive maneuvers to dodge the RPGs. BLAM! BOOM! BANG! The van flips off road."
Maybe when filming, they realize two rockets make more sense. Leave the implementation details to the experts.
However dialogue forms a large part of these scripts. Dialogue is engineered by writers, right down to the syllables. (Funny enough, AI screenwriters often forget syllables exist, and you can tell because they're difficult to actually speak)
What's the purpose of the spec? Instructions? To iron out risks and roadblocks? The document should aim for the bare minimum for that. What's your "dialogue" part - the thing that you need analysts to plan out precisely?
jcalloway_dev 19 hours ago [-]
Great experiment. The "implied context" problem is real and it kills projects.
One thing I'd push back on slightly: the 5 vs 127 framing makes this feel like a volume win for AI, but I think the actual insight is that AI externalizes the assumptions humans carry silently. That's the useful part.
What worked for us was using AI-generated specs not as a deliverable but as a conversation starter. You print the 127 points, sit with the client for 90 minutes, and the deletions become the spec. "We don't need multi-tenancy" is a real decision, not an oversight, once someone's forced to say it out loud.
To your questions directly:
1. Yes, reusable checklists for auth/RBAC/rate limits are underrated
2. 127 points is too many to hand a dev team, but perfect for a client workshop
3. Filter by "can we launch without this" — ruthlessly
Would love to see those prompts.
mackatsol 1 days ago [-]
Both are bad IMHO .. a senior human analyst who accepts a 2 sentence product description? What kind of weird reality is that coming from? I’ve had a client do that too, but it’s the analysts job to ask all the point follow-up questions so they end up with a proper requirements list. That ends up being multiple pages long.. feed that to the AI! I agree the AI as stated above has a bigger coverage, but it’s not doing a better job, it’s being just as lazy and adding a ton of filler to a lousy prompt. Yeah, this set me off. Great topic! Looking forward to reading the discussion. :)
allinonetools_ 1 days ago [-]
I have seen this play out on real projects. The missing edge cases are usually what cause delays, not the main features. Using AI as a checklist and then trimming it down with human judgment seems to work better than relying on assumptions alone.
A character description looks like this: "APOLLO CREED. Creed is twenty-eight years old. He is a tall, smooth-muscled Black with barely a scar on his light coffee-colored face."
It includes everything important for casting this guy. It doesn't say what he's wearing, hair style, things like a wedding ring. Once they cast an actor, the actor fills this up.
Fight scenes are designed to be as fast to read as the action. Poor writing is something like, "Terrorists B and C fire RPGs at the van. The van makes evasive maneuvers. After the third rocket, the van flips off road." It doesn't make the story nor the scene clearer.
The better script: "The van takes evasive maneuvers to dodge the RPGs. BLAM! BOOM! BANG! The van flips off road."
Maybe when filming, they realize two rockets make more sense. Leave the implementation details to the experts.
However dialogue forms a large part of these scripts. Dialogue is engineered by writers, right down to the syllables. (Funny enough, AI screenwriters often forget syllables exist, and you can tell because they're difficult to actually speak)
What's the purpose of the spec? Instructions? To iron out risks and roadblocks? The document should aim for the bare minimum for that. What's your "dialogue" part - the thing that you need analysts to plan out precisely?
One thing I'd push back on slightly: the 5 vs 127 framing makes this feel like a volume win for AI, but I think the actual insight is that AI externalizes the assumptions humans carry silently. That's the useful part.
What worked for us was using AI-generated specs not as a deliverable but as a conversation starter. You print the 127 points, sit with the client for 90 minutes, and the deletions become the spec. "We don't need multi-tenancy" is a real decision, not an oversight, once someone's forced to say it out loud.
To your questions directly: 1. Yes, reusable checklists for auth/RBAC/rate limits are underrated 2. 127 points is too many to hand a dev team, but perfect for a client workshop 3. Filter by "can we launch without this" — ruthlessly
Would love to see those prompts.