Plans are short and lack detail #649
Unanswered
mindplay-dk
asked this question in
Q&A
Replies: 2 comments 1 reply
-
|
I will revise the prompts. Jc , what model are you using? |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
I’m shipping a general solution tonight that can inject better planning
instructions, or for you to add custom ones
…On Sat, May 9, 2026 at 5:18 AM Rasmus Schultz ***@***.***> wrote:
Claude Sonnet 4.6 usually
—
Reply to this email directly, view it on GitHub
<#649 (reply in thread)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/ABXIUHLQDVR7QOLXPQOPK5L4Z3ZU5AVCNFSM6AAAAACYOHLNR6VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMOBWGEYDINY>
.
You are receiving this because you commented.Message ID:
***@***.***
com>
|
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
I worked through a fairly long session of discovery and planning with Pi, pulled in lots of documentation, made lots of technical and architectural decisions - and then invoked Plannotator, hoping to summarize everything, so I could clear my context and work from a clean slate.
Basically all it seems to capture is a short summary of high level requirements? There were so many technical details and nuances it just seemed to ignore.
I rolled back using
/treeand tried again with a prompt specifically asking for "very detailed" requirements and "capture everything" etc. - and it still just produces a very short 1-page document with the most important bulletpoints in very short form.I was expecting it to capture all the technical details I uncovered during discovery and planning - with all of this ignored, the agent would basically just need to read all of the code and documentation for two libraries a second time, likely ending up somewhere in the neighborhood of the same amount of context anyway, so then I'm not sure what the point would be?
Now, I didn't actually compare code generation with and without this plan, but I mean... if there is clearly not enough information in the plan to implement something, it can't know any better the second time which parts of documentation or source code read, etc. - so I'm assuming it would just need to read everything again.
I love the UI on this thing! It's very effective.
But anyone else feel like it does not really take instructions? It kind of seems to do what it wants.
I haven't read through the prompts, so I'm not sure why that is, but... anyone else? 😊
Beta Was this translation helpful? Give feedback.
All reactions