Pinned Tweet
HYBRD AI Toolkit:
1. Integ test your prompts with BAML
2. Optimize your prompts with the Prompt Optimizer
3. First prompt for a plan, then prompt to execute the plan
1. Integ test your prompts with BAML
2. Optimize your prompts with the Prompt Optimizer
3. First prompt for a plan, then prompt to execute the plan
good advice from @__ruiters:
GPT-5 isn’t broken. Your prompts are.
A lot of folks, myself included, expected GPT-5 to be fungible in the sense that you could drop it right into your existing workflows, and it would “just work.” But the depth of the GPT-5 prompt guide makes it clear: this is a major change; a major version bump, if you will.
Even the Cursor team, who lead pilot adoption for the new model, called it out:
> “GPT-5 is one of the most steerable models I've used. I've needed to be more explicit about what I was trying to accomplish for some tasks. Leaving things vague ended up with the model taking a different direction than I expected. When I was more specific, I was surprised by how smart the model was.”
Users’ frustration is Hyrum’s Law in action:
> With a sufficient number of users of an API, it doesn’t matter what you promise in the contract: all observable behaviors of your system will be depended on by somebody.
This basically means that OpenAI is big enough that now, no matter what changes they make, or how much better the models are, people will always complain because they have built their systems to depend on the behavior of the models at the time.
The OpenAI team has said it themselves: GPT-5 is extremely steerable. This is both feature and bug:
* It will do what you tell it to
* But you have to know what you want to it to do or at least articulate your intent better
At HYBRD, I’ve been using GPT-5 and finding it works quite well with appropriate prompting, and quite poorly without.
If you’re shipping on GPT-5:
1. Treat prompts like code: version, test, and review them. -- THIS IS THE MOST IMPORTANT ONE!
2. Read the prompting guide and understand how this model actually works (link in comments)
3. Run your prompts through the OpenAI Prompt Optimizer (link in comments)
4. Ask it to plan it’s approach before touching any code.
3
1
7
2,060
One of the main reasons to "read the code" that I haven't seen people talk about is lack of alignment between human and agent.
Inevitably, there are some edge cases that come up (especially in large code bases) that the human didn't clarify and the agent ends up having to make the choice.
Humans are terrible at explaining EXACTLY what they want, and at least for now, the best way to get around that is to keep a pulse on the code.
Inevitably, there are some edge cases that come up (especially in large code bases) that the human didn't clarify and the agent ends up having to make the choice.
Humans are terrible at explaining EXACTLY what they want, and at least for now, the best way to get around that is to keep a pulse on the code.
0
0
2
89
I love engineering twitter because you can just point your agents at tweets and have them build things
ty for the inspo @zeeg @0xblacklight
side: @bot should be able to access X?
ty for the inspo @zeeg @0xblacklight
side: @bot should be able to access X?
i wonder who is building the general-case solution for this:
1. deterministically run tests for code parts that changed (via AST?)
2. use jev to select test suites or something like that
3. declare globs in test files for files which when changed should result in the tests being run
0
0
3
157
ruiters retweeted
🚨🚨 talk drop 🚨🚨 - state of the software factory, live at @AgenticAIFdn in Amsterdam last week
don’t be the last one
youtu.be/ix1qQK1IvmA?is…
don’t be the last one
youtu.be/ix1qQK1IvmA?is…
2
17
208
27,179
ruiters retweeted
We see more and more people get rate limited by CI, alongside an increasing trend to move towards framework agnostic builds, to have more modularity and flexibility.
As verification becomes more and more important, the throughput of CI increases by multiple orders of magnitude.
integrate and own the entire stack, we can offset the cost to our customers.
As verification becomes more and more important, the throughput of CI increases by multiple orders of magnitude.
integrate and own the entire stack, we can offset the cost to our customers.
6
5
18
1,328
Can you believe engineers these days? Thinking for 6 seconds, deciding that they're "done" and taking a nap??
2
0
5
265
Text-to-speech is growing faster than ever but the tooling around AI audio is still super underwhelming.
But today celebrating a milestone - SpeechSDK just hit 10k/mo npm downloads!
This library is just my small open-source contribution to the audio ecosystem.
Backstory:
When a new TTS model is released (e.g. last week Gemini 3.8 Flash TTS, this week ElevenLabs v4), it's harder than it should be to swap models.
I should pick a model based on its output quality, not because I built business logic around certain provider behaviors.
The SpeechSDK is like @vercel's AI SDK for TTS, but specifically-tuned for audio and its edge cases.
For example:
Some models have native conversational dialogue. Other's don't. The SDK includes native multi-speaker stitching for unsupported models.
Some models include word timestamps in the response. Others return character-level granularity and some (cough @googleaidevs cough) none at all.
Some support emotive [audio tags] with square brackets. Others with <angled>. In some, this is unsupported and accidentally speak them (we don't want that!).
Some default to WAV or PCM, and others MP3.
Most models don't even normalize outputs to ensure all voices come out at the same volume.
As you dig into each more, you start to find little provider-specific quirks and edge-cases that your business logic depended on, making that TTS model change not an easy swap. And since AI still can't "qualitatively hear" well, some of these bugs often go unnoticed.
All these audio-related things are solved in the SpeechSDK. Try every new model with a single line string change and it just works.
It's open-source, Apache-2.0, and provider agnostic.
Github: github.com/Jellypod-Inc/s…
2
0
5
402
HYBRD rn:
We’re blown away by the response to the @Muse Connector Platform — with over 2,000 submissions in just a few days since launch. We believe this is the beginning of a new era for developers big and small, and we’re so grateful for the response from the community to date.
For you can learn more about the platform and submit your connectors here: muse.ai/platform
0
1
4
289
models reasoning about their own tokens is weird...
"Evaluating testing budget
I’m trying to figure out my token budget for a task where I have 14k tokens but only 13 left. The task is complex!"
"Evaluating testing budget
I’m trying to figure out my token budget for a task where I have 14k tokens but only 13 left. The task is complex!"
2
0
2
129
Queue the next round of “OMG <insert new model> just built this <insert 3d rendering of thing>” tweets
0
0
3
267
Maybe an obvious take, but agents paying for things for you is going to similar to self driving cars or getting into strangers cars and having them drive you somewhere
1
0
0
82