<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: DevinDespair</title>
    <description>The latest articles on DEV Community by DevinDespair (@devindespair).</description>
    <link>https://dev.to/devindespair</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4250478%2F87b1e7ad-8a81-498e-b279-c3a81d887109.jpg</url>
      <title>DEV Community: DevinDespair</title>
      <link>https://dev.to/devindespair</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/devindespair"/>
    <language>en</language>
    <item>
      <title>Why we built Cinder</title>
      <dc:creator>DevinDespair</dc:creator>
      <pubDate>Tue, 17 Nov 2026 19:25:41 +0000</pubDate>
      <link>https://dev.to/devindespair/why-we-built-cinder-3dc0</link>
      <guid>https://dev.to/devindespair/why-we-built-cinder-3dc0</guid>
      <description>&lt;p&gt;The pitch behind most "talk to your app" tools is the same: wrap an LLM&lt;br&gt;
around your UI, send every command to the cloud, let the model figure&lt;br&gt;
out what the user meant. It works, demo-well, and it's expensive and&lt;br&gt;
slow for the thing it's actually used for most: a user asking their own&lt;br&gt;
app to do something the app's developer already knows how to do.&lt;/p&gt;

&lt;p&gt;"Open the reports page." "Clear the crop filter." "Show soybean prices&lt;br&gt;
in Punjab." None of these need a frontier model to understand, the&lt;br&gt;
developer who built the reports page, the crop filter, and the price&lt;br&gt;
lookup already knows exactly what those commands mean. The problem&lt;br&gt;
isn't that the app doesn't understand the user. It's that there's no&lt;br&gt;
structured way for the developer to tell the runtime what they already&lt;br&gt;
know.&lt;/p&gt;

&lt;h2&gt;
  
  
  Semantic tags, not prompts
&lt;/h2&gt;

&lt;p&gt;Cinder's bet is that a handful of plain HTML attributes, cinder-route,&lt;br&gt;
cinder-action, cinder-field, cinder-filter, cinder-output, are a better&lt;br&gt;
interface for this than a prompt. A tag is unambiguous, versioned with&lt;br&gt;
your code, and costs nothing to evaluate. A prompt is a guess,&lt;br&gt;
re-evaluated on every request, with real latency and a real bill&lt;br&gt;
attached.&lt;/p&gt;

&lt;p&gt;So tier 1 is a deterministic graph: tag your real routes/contexts/&lt;br&gt;
actions/fields, Cinder scans the DOM once, builds a scoped graph, and&lt;br&gt;
resolves most real commands against that graph directly, no model, no&lt;br&gt;
network call, no bill. We found, building and testing this against real&lt;br&gt;
apps, that the honest majority of what users actually ask for falls&lt;br&gt;
into this bucket once an app is reasonably tagged.&lt;/p&gt;

&lt;h2&gt;
  
  
  The AI tiers exist for the honest remainder
&lt;/h2&gt;

&lt;p&gt;Tier 2 (@usecinder/embeddings) catches typos and paraphrases of things&lt;br&gt;
that are already tagged, still entirely client-side, still free,&lt;br&gt;
because recovering "clera the croup filter" doesn't require a cloud&lt;br&gt;
round-trip either.&lt;/p&gt;

&lt;p&gt;Tier 3, Relay, is what's left: requests the deterministic graph and the&lt;br&gt;
local model genuinely can't resolve, usually open-ended or analytical&lt;br&gt;
ones. This is the one tier that costs real money, and we built it to be&lt;br&gt;
billed like what it actually is: a prepaid wallet metered against real&lt;br&gt;
token usage, never a flat fee for a feature most commands never touch.&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest tradeoff
&lt;/h2&gt;

&lt;p&gt;This only pays off if you're willing to tag your app. Cinder is not&lt;br&gt;
going to understand an untagged page the way a general-purpose web&lt;br&gt;
agent might try to. We think that's the right tradeoff for an app you&lt;br&gt;
actually control and ship, the alternative is paying cloud-AI latency&lt;br&gt;
and cost on every single command, for an app whose structure you&lt;br&gt;
already know perfectly well.&lt;/p&gt;

&lt;p&gt;If that tradeoff makes sense for what you're building, the install is&lt;br&gt;
one line: npm install usecinder. Full docs at &lt;a href="https://usecinder.dev" rel="noopener noreferrer"&gt;https://usecinder.dev&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>softwaredevelopment</category>
      <category>webdev</category>
    </item>
  </channel>
</rss>
