The latest articles on DEV Community by Tanishq Agarwal (@tanishq286). https://dev.to/tanishq286 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%2F3977454%2F63556c9c-e5c8-497e-a393-7279a5e053e4.png https://dev.to/tanishq286 en Tanishq Agarwal Fri, 13 Nov 2026 16:52:18 +0000 https://dev.to/tanishq286/cut-my-always-loaded-claude-code-context-by-575-7337-3121-tokens-and-built-a-2oba https://dev.to/tanishq286/cut-my-always-loaded-claude-code-context-by-575-7337-3121-tokens-and-built-a-2oba <p>After building dozens of custom skills and subagents in Claude Code, two problems kept biting me:</p> <ol> <li>Always-loaded skill definitions were burning 7,300+ tokens before I even typed a prompt.</li> <li>A skill that worked great on Opus would quietly degrade on Sonnet or Haiku—and pinned model IDs in templates kept going stale.</li> </ol> <p>So I built <strong>tq-forge</strong> (MIT) — an open-source CLI and Claude Code plugin that scaffolds, learns, scores, compresses, and promotes skills and 5-file subagents.</p> <h3> What it actually does under the hood </h3> <ul> <li> <strong>Lossless Context Optimizer (<code>/savings</code>):</strong> Compresses always-loaded skill context without losing MCP server availability. On our test-suite benchmark fixture, it cut always-loaded tokens from <strong>7,337 down to 3,121 (-57.5%)</strong>, and is 100% reversible with <code>restore</code>.</li> <li> <strong>0–100 Structural Rubric + Model-Compatibility Lint:</strong> Runs 320 deterministic scorer assertions and checks model compatibility before promoting any artifact from sandbox to production (<code>/tq-forge-test</code>, <code>/tq-forge-promote</code>).</li> <li> <strong>Forge from Intent or Auto-Learn (<code>/tq-forge-skill</code>, <code>/tq-forge-agent</code>, <code>/tq-forge-learn</code>):</strong> Generate a <code>SKILL.md</code> or 5-file subagent from a one-line prompt—or point <code>/tq-forge-learn</code> at a doc page, browser workflow, or technical thread to deduplicate links (<code>linkdedupe</code>) and distill repeatable steps into a scored skill.</li> <li> <strong>Weekly Model Sync:</strong> Automated GitHub workflow that detects new Claude model IDs, updates templates, runs 165 unit tests, and opens a PR so agents don't drift onto retired models.</li> </ul> <h3> Install &amp; Try </h3> <ul> <li>Claude Code Plugin Marketplace: <code>/plugin marketplace add tanishq286/tq-forge</code> </li> <li>npm CLI: <code>npx tq-forge install</code> </li> <li>GitHub (MIT): <a href="https://github.com/tanishq286/tq-forge" rel="noopener noreferrer">https://github.com/tanishq286/tq-forge</a> </li> </ul> <p>Happy to share the scoring rubric or optimizer diff if anyone wants to inspect how the token compression works!</p> <p>If <code>tq-forge</code> saves tokens or catches model drift in your Claude Code setup, <strong>drop a star on GitHub</strong> (<code>https://github.com/tanishq286/tq-forge</code>) — it helps other Claude Code builders discover the project and directly supports our open-source cross-model eval roadmap!</p> opensource ai claudecode productivity Tanishq Agarwal Wed, 10 Jun 2026 10:23:14 +0000 https://dev.to/tanishq286/i-built-a-token-free-deterministic-scorer-for-ai-outputs-and-why-most-evals-are-broken-na7 https://dev.to/tanishq286/i-built-a-token-free-deterministic-scorer-for-ai-outputs-and-why-most-evals-are-broken-na7 <p>Liquid syntax error: Unknown tag 'endraw'</p> ai claudecode devtools opensource