... nobody covered, which is worse, because it changes a line of code a lot of us shipped this year without thinking about it. If your product has a "paste a URL" box, a link unfurler, an enrichment step, a competitor-monitoring job, or anything that fetches a page and hands it to a model, this is about you. What was actually measured. The paper is "Indirect Prompt Injection in the Wild" by Khodayari, Zhang, Acharya and Pellegrino, out of CISPA ... ... hidden - the comment tags, the display:none, the fact that this text sits in a meta field. When you strip to plain text with something like BeautifulSoup, you delete that evidence and flatten a concealed instruction into ordinary prose

... minute check for your own repository, and it is not the one you think. Take a number out of your infrastructure config. Worker count, pool size, connection cap, a timeout. Then search your source for the prose that argues from it not the number, the sentences. The ones that say: because we have N of these, this limit is X. Mine: the app server's HTTP pool is six concurrent workers. That number is well guarded. There is a unit ... ... because somebody did check, in July. Two things kept it alive. The background jobs pool really is two, so two is a live number in that same config file and nothing looks odd next to it. And nothing tests prose

... scenes you re actually writing is another challenge. Skeleton Draft grew out of how I work through that relationship between story and structure. It helps you shape your story through acts, chapters, scenes, and beats, then write it as prose, a screenplay, or both side by side. You can move between the big picture and the writing without losing your place in either. I m bringing Skeleton Draft to the Product Hunt community and would love ...

... burned by this. My site says things like "152 label formats" and "114 countries" on hundreds of pages. When the real count changed, prose across dozens of pages kept the old number for weeks. What I do now: The counts live in one constants file, and pages read from it instead of hard-coding. A prebuild script scans every generated page and fails the deploy if any page states a format count, country count or plan limit that disagrees with ...

... name, so a repricing withdraws it from the page instead of quietly making the page wrong. The test reads the rendered body rather than the method behind it, because the next hard coded price written into the prose would never go near that method. Two questions, and I would rather have the second one answered. What did you find your cached pages were varying by? And has anyone got a check that catches a value going stale inside a cache ...

... able to read what the new one writes. It's invisible on day one because you're the only consumer of your own JSON. We only call models with a strict JSON schema and never by parsing prose out of a response, and strict mode has edges you inherit as product constraints. It rejects minItems, so "at least ten of these" is something you handle in code, not something the schema can promise. The other thing I'd want ...

... Then I started a run at 01:12 and it wrote that day's report straight into yesterday's file. The previous day's numbers were gone. They exist now only because I had also typed them into a prose log by hand. What I had wrong was the cost model. I kept pricing that line at two minutes, so it lost every morning to whatever was actually bleeding. The real price was two minutes multiplied by the odds ...

... sessions, where do you keep the authoritative scope? In the issue, a state file, the orchestrator or somewhere else? How do you notice when the work has drifted? Comment from Alex Gidirim(@alex_gidirim): @siarheihamanovich 's shift from prose rules to executable tests is the part that actually holds up. A spec sitting next to the agent only works if someone rereads it every session, and that discipline erodes exactly when you need it most. Turning the rules that keep ... ... stop there. Orchestrator memory and chat history both quietly rewrite the goal, but git blame doesn't. Comment from Siarhei Hamanovich(@siarheihamanovich): What changed things for me was moving the small number of rules that actually hurt out of prose

... Fixed Mailbox chat works again on gpt-5.6 models. Point the OpenAI-compatible provider at OpenAI with a gpt-5.6-* model and every single round that carried tools came back as a 400 and a wall of English error prose in the chat bubble. Retrying showed you the same wall. Very helpful. ( #40 ) The twist: Skim never sent reasoning_effort at all. The model reasons at medium on its own, then refuses its own default alongside our tools so nothing ...

... worth looking at what they are actually testing, because it is almost never style. Cadence, tidy paragraphs, punctuation habits, those are proxies, and every one of them comes out of a single prompt, so a founder who tightens their prose gets accused more often. What a prompt cannot produce is a claim that could turn out to be wrong. Generated text is unfalsifiable because nothing behind it happened, so the only signal that survives is detail specific enough to check ...

... text back, maybe a code block. We wanted Kopai's agents to answer the way a good analyst would with the actual chart, table, or calculator, not a paragraph describing one. So when an agent decides a visual beats prose, it writes real render-able code and it renders live in the chat inline, not a separate canvas. It runs in a sandboxed iframe with a null origin, so the generated code can't touch your cookies or session, even ... ... tool that should've just been a table or a chart? Comment from Raffay Sajjad(@raffay_sajjad): Mine is expense category breakdowns, every finance AI I've tried gives me "Category A is 40%, Category B is 25%..." in prose

... into a machine, one day. They are mostly racing on the same three things: accuracy, latency, and whether it runs locally. All real, all table stakes within a year. Here is what I think the category is missing. Dictating prose and dictating to an AI are different problems, and almost everything is built for the first one. When a human reads the output, polish matters. Strip the filler, fix the punctuation, tidy the sentence. Every one of these ships some ...

... tester the artifact and the original spec and nothing else. No builder reasoning, no summary, no "I have verified this works". Then define success before the build starts as something that exits zero or non zero, never as prose. If the criterion is prose

... needs less explanation than you think, because the right format is a table rather than a paragraph. Four columns: what connects, where to, what it contains, can you turn it off. One row per connection. Nobody reads prose about networking, but everyone reads four rows, and a table is also the only version that stays honest, because adding a connection later means adding a row and you will notice you are doing it. Comment from Dale Mooney(@dalemooney): Rabnoor ... ... distinction is the right one. I'd add the mechanism that makes it decisive inside a company, rather than just principled. If you're the person who signs a tool off, cloud versus

... nothing, the reviewer's normal experience becomes an empty screen, and empty screens train people to expect empty. Whether the one that is not empty gets read properly on day forty is an open question. It still beats the alternative, because a difference against a background of silence is about as salient as anything gets. And yes to the small pile. Most of what gets called judgement is work nobody has bothered to make mechanical yet. Comment from Aadil Ghani ... ... structured field for a source, so provenance in our asks is whatever the agent chose to say about itself, which is precisely the narrated kind I just argued against. A field that resolves to an artifact instead of prose

... different file. The split that held up for us on the support side: can the constraint fail loudly, or can it only be read? Anything you can turn into a check keeps working when nobody remembers it. Prose is a reminder, and a reminder loses to whatever was said most recently. Earliest sign for me was re-explaining the same constraint twice inside one session. Not across weeks, within one sitting. That is the moment the model stopped carrying ...

... normal conversation. Good punctuation and fast corrections would probably matter more to me than extra AI features. Comment from Rabnoor Singh(@rabnoor_s): Yes, and I think dictating to an AI is a genuinely different problem from dictating prose. Most apps are built for the second one. Two things that fall out of that. The cleanup pass is the wrong default for prompts. Stripping filler and fixing punctuation is right when a human reads the output. When a model reads ...

... both. The failure was not a wrong answer. It was a partial check summarised as a complete one. "RLS is enabled" is a fact that sounds like a conclusion. What changed: I stopped accepting any security claim in prose. If the agent says something is protected, it has to produce a test that fails when the protection is removed. If it cannot, the claim does not exist. That habit stuck, probably because it is mechanical rather than a matter ...

... inputs, sensible defaults, and a cap on how much context a request can pull cut spend before a token is ever priced. Output usually costs more than input, so anywhere you can return a short structured answer instead of prose is a saving the customer never notices. And watch the median user, not the mean. A few power users drag the average up and panic you into changes that punish the 80 percent who cost almost nothing. Optimise ...

... manually score) and re-run them after any prompt change. Cheap, catches regressions fast. 2. Structured output over freeform forcing JSON schemas with explicit fields (score, reasoning, flagged_items) makes drift way more visible than letting the model write prose. If a field goes missing or the reasoning gets vague, that's a signal before a user even complains. 3. Boundary tests, not just happy-path tests for the riskier tools (anything insurance/legal-adjacent) I specifically test adversarial inputs, people trying ...

... much time, exactly, does kubelet give it? A default that shouldn t be a default The answer is in the Kubernetes documentation , but it s easy to miss because it s in the table of defaults rather than the prose: timeoutSeconds defaults to 1 . failureThreshold defaults to 3. periodSeconds defaults to 10. Translated: kubelet calls /health, waits up to 1 second for a response, and if it doesn t get one in that window the probe is failed. 3 failures ...

... evolves, the governance evolves with it rather than falling behind. Would love to hear more about where the sync breaks down for your team specifically. Comment from Ahmad(@blah_mad): This is the part I would want outside the prose file: a small contract that can fail when repo behavior drifts. Otherwise AGENTS.md becomes a policy doc with no test attached. Comment from Ahmad(@blah_mad): @achille82    Small practical note from the agent-builder side: before a tool ... ... canonical place the agent always reads but for the handful of operations that could cause real damage, enforce it with something that can't be ignored, not just described. Curious to see where Ota lands on that doc-vs

... doorbell. It rings the same whether it's a package or a fire. That's basically the gap Pushary fills, same hook idea you've already wired, but the event carries a real handoff (done vs blocked vs

... system contradicts something you wrote in chapter four, and now you need a lore bible. So you open a wiki tool. Then you need an outline, so you grab a plotting app. Then you want to clean up your prose, so you sign up for a grammar checker. Then you want to format your manuscript for publishing, so there s another tool. Before you know it, you re running five or six different apps just to write a single book ... ... finally had enough of Microsoft Word. I was writing a fantasy series with deep worldbuilding, hundreds of characters, multiple magic systems, interconnected timelines and Word was buckling under the weight of it. But when I looked at the alternatives

... Obsidian. Logseq. Notion. Beautiful graphs. Lots of [[backlinks]]. And I still forgot what I read last week. Why? Because linking isn't learning. Encoding is. Cognir enforces encoding: Read Annotate with your thought (no skipping) Graph appears Compile to prose or report It's not a database. It's a thinking instrument. If you've ever stared at your own PKM and thought "this feels like a filing cabinet for stuff I don't actually know" try this. Full local ...