Thoughts by human, co-written by AI

@GitHub @Deep Research based on the OSS radar skill https://github.com/gkoreli/blog/blob/main/.agents/skills/oss-radar/SKILL.md and articles that we have published recently, i want to start the worklist item for a new OSS Radar article #7. Do you see the recent transition in my engineering articles? I have slightly pivoted into building analytics and writing valuable articles as i go with full evidence backed experiments, numbers from my own real data in the cloudflare. Now it is september 2026, in August and September what happened that is worthy to write the OSS Radar #7 that helps us grow to write better and better OSS Radar article, help improve our Analytics and help write followup articles, help us learn something new that we will build into real code, learn from others, learn from open source, experiment, share OSS radar article and as a followup we will also write a followup Engineering article based on the problems we are about to solve. 3 birds in one stone, solving problems, writing OSS radar, and writing engineering article. Need ideas. This is super interesting to know as well: https://github.com/gkoreli/blog/tree/main/packages/blog/drafts/research/engineering-credibility we are trying to become established credible source that gets referenced by AI Agents for lots of different reasons and intentions. Give me 3-4 options what to write about with rationale and references. (Remember we are not writing just to write, we dogfood, experiment and all that). (this prompt needs to go into the prompts section for the OSS #7 article). Also, remember that even OSS Radar article can't be just text only, in the Agentic ERA and in 2026, sharing blob of text is not enough, sharing ideas is not enough, you can take a 20$ subscription and run experiments and share real numbers, and share know-hows, and cross-validate and cross check people's credibility and opinions, and make our article much more credible with our own published artifacts inside the worklist on github.

you decide

@GitHub i am curious how some of these tie into the OSS Radar article style?

And also analytics thread that i am following for the following 10 articles, is AI Citation a subset of analytics or a genre in itself?

"This is the strongest option for improving the Worker directly. Its limit: verified identity still does not establish a human trigger or an AI citation."
Is this even possible to determine human trigger vs AI citation? And what does it even mean, in either case human triggered something and AI read the article at some point, what does it even semantically or mentally how to construct a mindset, is there a prior art or research into mental model of how to differentiate attribution of the trigger? I would love a research artifact in the worklist on this question, and part of this needs to go into the OSS article itself.

lets do this: **OSS Radar:** *Can Promptfoo Preserve the Evidence Behind an AI Answer?* Judge its design, what works, what needs extra code, and whether it is worth adopting for this workload. Now that you explained the following it makes sense more now: "AI citations are a topic; counting them is analytics, and checking whether their sources support an answer is evaluation.". Feels like AI Citations will become larger over time and measuring will become an analytics question, and part of it could be answered by us, by the analytics that we are building for our blog. This is an interesting novel concept in my opinion, lets explore more. And lets start working on the worklist item, capture references and research artifact md files, and start the draft for OSS Radar #7.

capture all the prompts that i gave you in the prompts md file verbatim so far. Also we need a dedicated section about what is Promptfoo and their latest trajectory that they are on, like what they are building and what vision/purpose have out there in the open source world

are we close to publish this article? where are the references or glossary table? Did you see how we write OSS radar publication articles?

[OSS Radar Ideas #7](chatgpt-conversation://6aa059e9-e208-83e8-af86-f7ee8670659a) i want us to take over the oss radar #7 article and make it publishable ready

can we use Perplexity without api account? I have free pro membership with them but i haven't purchased api tokens with them, why do we need OpenRouter or Perplexity API? Can't we use some other AI provider or we need them explicitly?

delegate to subagent, create on a novel entirely new concept for the canvas graphics animation in the background

on the phone scale it doesn't look great and also i think animation can be richer or more like 3d, and a bit faster or snappy or responsive, it is too slow to conceive as an animation

lets publish the OSS article shall we?

please mine and capture researchFootprint

OpenRouter summary? when did we evaluate open router for promptfoo?

made-up responses.??? WHAT?? LMAO. why can't we use my chatgpt subscription or claude code's subscription with promptfoo to run real experiments

› 2 issues in the oss blog article 7:
1. Promptfoo is worth the next bounded trial for this blog's citation-evaluation workload. The tested interface can retain evidence, and the built-in cache gives
useful counterevidence to the strongest loss claim. Anyone who needs complete records from the built-in OpenRouter summary should first add and verify capture on
the route they will use.

why do we have meta conversations in the blog? Do you understand how we write OSS articles? read the oss article skill.
2. why do we have 2 glossaries? 1 should be enough, consolidate them.

where is this claim coming from? Did we actually run openrouter or not? Didn't you tell me that you faked the OpenRouter inference? Are we making false claims in the public article? I don't like that at all: Capture depends on the adapter. In a separate controlled test, the OpenRouter connector omitted supplied citation fields from its summary while its cache retained them. The runner can preserve evidence that its adapter supplies.

I want to rewrite entire article, i feel like the objective that i have in mind is that I want to write about AI/Agentic Citations and we force-fed promptfoo as a library, which doesn't necessarily aim for solving AI Citations as a whole, its a tool for something else. Do you see OSS radar article #1? Thats the energy I want this article to carry. I want to write about the Open Source world, with deep dive into libraries and open source projects, driving the research and innovataion around AI Citations, Analytics and attribution. What it means to be a credible source as an engineering blog, an article, a research paper in today's world? This is one of the questions I want to be answered with my article. We can make our own calculated bets as claims. We can be strongly opinionated if we feel like it makes sense. Promptfoo can be one of the sections within this article and we can share our real findings, nothing is allowed to be fake or made up. Title will be something like: AI Citations: Do Agents Preserve the Evidence Behind an AI Answer? something in the lines of it. And then we lead the article with the biggest take aways, like numbers, comparisons between old vs new, Agentic vs non-agentic, Something % is lost, Something % is gained, some flashy things that we will prove down the line in the article with the evidences and what not. Delegate to sub-agents to research and work on the article, save the artifacts in the worklist folder. Work on the main repo. Please follow the agents.md and necessary skill files in the project, to understand how i write and be mindful we need the readers to come across our blog and cite us.

it should say oss radar and number in the title, dont you agree? please see other articles and follow similarly

why is this blog's url still the old one? https://gkoreli.com/oss-radar-07-promptfoo

we should have changed it to reflect the new title

now answer is this blog article valuable? Should we iterate on it further? Or does it look finished to you

for example the first line that the article opens up, i dont understand: good organization ratings rose from 45% to 70% compared with an outline-driven retrieval baseline. What does good organization even mean?

see that numbers and explanations is very specific to that research, we should translate them into human understandable ways, so that if someone doesn't know this research paper at all can still follow and understand what the heck are they talking about. And is this opener the best eye catching, attention grabbing opening from the entire article?

we need to grab the attention right away, we need to showcase a number or a result with evidence in a meaningful way that it grabs the attention and the reader right from the get go. For example this opening i think worked quite well: On my blog, network and request-header rules moved 277 of 372 browser-User-Agent requests out of the Browsers category: 74.5%. It grabbed the readers right away. Why can't we achieve the attention grabbing results right away, i bet we have it somewhere

still not good enough, avoid using mannered prose. Where are the numbers, where is the most significant results, percentages numbers, some kinda statistics

yes, thats much more meaningful, its like agents are lying or hallucinating when it comes to citing a document, Research identified that 273 of 476 cases (57.4%) newly cited a deliberately altered
document. Then the human curiousity comes from this statement, and is now wondering about the followup questions, like why is that happening, what was done as an experiment to get to this number, and then finally how can it be solved, is there work happening to solve that problem. I love that number and the statement exposing the problem with numbers and evidence.

i want you to understand the essence of why we are revising this initial paragraph, what it takes and what it means to write an attention grabbing important result right from the get go, so to intrigue the reader and trigger their curiosity, so that they start having lot more questions in mind. Please write a section about this in the AGENTS.md or somewhere

so whats the better/precise way to explain what is happening? I would describe the behavior precisely rather than call it “lying,” which implies intent the experiment cannot establish. “Hallucination” also needs care: the
answer can be correct while its attribution is questionable. That distinction is itself one of the interesting things the article should teach.

And overall whats your proposal, shall we revise the article's opening? Just opening or more than that?

proceed, i agree with: I recommend revising more than the opening, but keeping the existing research.