Andrei's Email builder #382
Conversation
…individually. This version is incomplete, as the backend logic has not yet been updated.
… to planning out the rest of the functionality. Prompt changes and which data gets passed should be the last things before the new system is operational.
… limits and a myriad of visual bugs, it seems to be working correctly on my end. The search functionality is so far only used for researching the target of the email. The code is messy, bug filled and needs lots of cleaning.
…write, mostly type declarations, checking that the errors still present are, in fact, result of outdated check files.
✅ Deploy Preview for pauseai ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
|
Huh, I thought we already had a PR but thanks |
|
Does anyone know any better ways to get around the netlify 30s timeout for serverless functions other than background functions? Those have a 15 minute timeout, but require a paid subscription and don't directly pass outputs. |
|
Hi! My user feedback
|
Reset all button now calls "clear" instead of "runTest" clear() now force updates the arrays so the contents display immediately. :wq
|
Copying from https://discord.com/channels/1100491867675709580/1222902730083799120/1386728025625727018 - Pro gets us background serverless functions. Context for short-term spend option. It's quite plausible the builder api should move outside of Netlify altogether, but short-term that's non-trivial and even long-term one deployment system / web framework has upside. Andrei had other options to squeeze below timeouts too, basically it just drops the research quality a bit. $20 for a trial to cover flexibility about this during the PauseCon opportunity seems a no brainer. But Joep owns the account - are there other complications? |
|
First impressions: The helper texts that say "Question x" make me think I may need to write a question there ? But the headers for each textbox don't seem to imply that. I think those headers could be a lot clearer about what I'm supposed to write there. I think the tabs and the rest of the buttons should be really differentiated from each other, They are really confusing at this time. I would expect the second row to be other tabs or sub-tabs. I don't really understand what the write button is supposed to do. I could continue but I think those are the most important points probably. Maybe in a call it would be faster to talk about the tool if you want even more feedback. |
|
It's not harsh. This is a power tool prototype. A real workflow gets a
better UX, with friendly guidance.
Even then, it will still be the case that some will often or always DIY.
The idea is to motivate and unblock a cold start (write persuasively to
people I guess at, how can I possibly) and to give those who value a
toolset more to try.
…On Tue, 24 Jun 2025, 23:34 Pato-desu, ***@***.***> wrote:
*Pato-desu* left a comment (PauseAI/pauseai-website#382)
<#382 (comment)>
First impressions:
Sorry if this is too harsh but I feel using the tool would be more
complicated than writing the email myself.
There are lots of buttons, lots of boxes on each tab, a bunch of text.
Don't really know where to begin and how everything connects to each other
and if the tabs communicate with each other.
The helper texts that say "Question x" make me think I may need to write a
question there ? But the headers for each textbox don't seem to imply that.
I think those headers could be a lot clearer about what I'm supposed to
write there.
I think the tabs and the rest of the buttons should be really
differentiated from each other, They are really confusing at this time. I
would expect the second row to be other tabs or sub-tabs. I don't really
understand what the write button is supposed to do.
I could continue but I think those are the most important points probably.
Maybe in a call it would be faster to talk about the tool if you want even
more feedback.
—
Reply to this email directly, view it on GitHub
<#382 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAAXXDJSNRDQY4SDPKYJD6L3FHHABAVCNFSM6AAAAAB77V757WVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTAMBSGA3DMNBSGI>
.
You are receiving this because you commented.Message ID:
***@***.***>
|
|
My proposal is that most of the prompts (including the auto-generated ones)
are hidden from the user (at least by default.)
A very common use case for the tool (including one we would like to
exercise at PauseCon) is that by default the user isn't doing much more
than say where they live (tool researches who to write to and their
background - most users won't edit the findings) and then choosing a "Ask
my MP to support our Deep Mind demand" option (producing a template for the
"ask" field - most users won't edit it.)
It writes a letter to that target, on that topic.
User should be able to continue the session (ask the LLM to change the
draft in various ways) or otherwise copy out the text and make their own
edits before sending.
Entering / editing prompt auto-fills should at least not be a prominent
feature.
…On Wed, 25 Jun 2025, 10:22 Anthony Bailey, ***@***.***> wrote:
It's not harsh. This is a power tool prototype. A real workflow gets a
better UX, with friendly guidance.
Even then, it will still be the case that some will often or always DIY.
The idea is to motivate and unblock a cold start (write persuasively to
people I guess at, how can I possibly) and to give those who value a
toolset more to try.
On Tue, 24 Jun 2025, 23:34 Pato-desu, ***@***.***> wrote:
> *Pato-desu* left a comment (PauseAI/pauseai-website#382)
> <#382 (comment)>
>
> First impressions:
> Sorry if this is too harsh but I feel using the tool would be more
> complicated than writing the email myself.
> There are lots of buttons, lots of boxes on each tab, a bunch of text.
> Don't really know where to begin and how everything connects to each other
> and if the tabs communicate with each other.
>
> The helper texts that say "Question x" make me think I may need to write
> a question there ? But the headers for each textbox don't seem to imply
> that. I think those headers could be a lot clearer about what I'm supposed
> to write there.
>
> I think the tabs and the rest of the buttons should be really
> differentiated from each other, They are really confusing at this time. I
> would expect the second row to be other tabs or sub-tabs. I don't really
> understand what the write button is supposed to do.
>
> I could continue but I think those are the most important points
> probably. Maybe in a call it would be faster to talk about the tool if you
> want even more feedback.
>
> —
> Reply to this email directly, view it on GitHub
> <#382 (comment)>,
> or unsubscribe
> <https://github.com/notifications/unsubscribe-auth/AAAXXDJSNRDQY4SDPKYJD6L3FHHABAVCNFSM6AAAAAB77V757WVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTAMBSGA3DMNBSGI>
> .
> You are receiving this because you commented.Message ID:
> ***@***.***>
>
|
|
Are you planning on finishing it before PauseCon? Maybe the whole Software team (+ me) should be working on this, then? @anthonybailey |
|
I'm going to remote pair with Andrei tomorrow, then demo result to Ella and fix whatever we have up however is best for the PauseCon session where folk play with it to send DeepMind stuff to representatives. If particular work (weird tech investigations) separate out then unblocking contributions are welcome, but for the most part I do not think this parallelizes beyond two. |
…uce the amount of buttons.
|
I'm going to remote pair with Andrei tomorrow, then demo result to Ella and fix whatever we have up however is best for the PauseCon session where folk play with it to send DeepMind stuff to representatives. If particular work (weird tech investigations) separate out then unblocking contributions are welcome, but for the most part I do t think this parallelizes beyond two. |
|
I've been stupid sleepy, but did some work with Andrei yesterday/today and there was a lot more from him. I expect to code and make the yes/no call on one key unknown technical issue (background functions for longer LLM think time greater robustness) tonight. Even if we get unlucky on that unknown, Andrei raised a possibility I had not considered and it is a great idea: if we end up only having the highest quality version of the writer running on a developer machine rather than the public website, we could serve still the website page that way at PauseCon - at minimum for a demo, and (idea occurring as I write) possibly across the local network so everyone could use it.) And we're making various steps in the direction of simple UI for the tool in general and the PauseCon ask in particular. |
Update for Andrei - Testing Session Results (2025-06-27)Hi Andrei! Here's what I discovered during overnight testing: Key Findings1. Rate Limit Death Spiral 🚨
2. Billing Discovery 💰
3. Usage Logging Implemented 📊
4. Model Comparison Plan 🔄
What I Didn't Test
Recommendation for DemoWithout model switching or rate limit improvements, we should:
The AIXR content degradation issue remains (firstCut removes key terms), but that's a separate fix. Let me know if you want the full testing details or have questions! |
- Implement comprehensive usage logging with rate limit tracking - Opt-in via write-usage.log file existence - Capture token usage, rate limits, duration, and errors - Clean wrapper pattern keeps server code uncluttered - Logs both successful API calls and rate limit errors - No performance impact when logging disabled This helps monitor API usage patterns and debug rate limit issues during conference demos and multi-user testing. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <[email protected]>
Update: Logging Implementation CommittedHi Andrei - I've committed the usage logging implementation to help track API usage during demos. To get the changes into your fork:Since I pushed to PauseAI/pauseai-website (not your fork), you'll need to pull the changes: # Add PauseAI repo as upstream if you haven't already
git remote add upstream https://github.com/PauseAI/pauseai-website.git
# Fetch and merge the logging changes
git fetch upstream locally-working-snapshot
git merge upstream/locally-working-snapshot
# Push to your fork to update this PR
git push origin locally-working-snapshotWhat's included:
The commit is: 2586555 on PauseAI/pauseai-website |
Final Update: Max Subscription ≠ API AccessResolution: Just got clarity from Anthropic support via Fin: The Answer
What This Means for Demo
Technical Impact
The system works technically, but web search + sustained multi-user load requires careful cost/rate limit management. Current StatusAPI functioning normally once we use the right billing model. Ready for demo with proper expectations about web search limitations. |
… Switched Write-AI to Claude-3.5-Haiku.
…em prompt. Introduced a revise email form for making edits.
|
Andrei in particular: I am making UX changes on to improve workflow on dev, particularly for PauseCon DeepMind usecase (but it's a good shift wider than that.) Code will go on the locally-working branch, avoid big refractors there. Changes to the background functions branch not blocked, but won't go there until subsequent round of productionization. Production has a key enabled and I also have a local key of my own. Current cost even without b"e as quick as you can" prompting is reasonable. Quality acceptable but improvable. LP |
@AndreiPauseAI created it, I just opened this PR to have a place to track progress and discuss things.
Preview