The latest articles on DEV Community by awesome fred (@awesome_fredtv_c0c3). https://dev.to/awesome_fredtv_c0c3 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%2F4168095%2F1a77aa92-f48a-4c97-af61-56c1f1d44d83.png https://dev.to/awesome_fredtv_c0c3 en awesome fred Wed, 07 Oct 2026 08:55:04 +0000 https://dev.to/awesome_fredtv_c0c3/token-saving-tools-made-my-claude-code-usage-go-up-heres-the-measurement-edk https://dev.to/awesome_fredtv_c0c3/token-saving-tools-made-my-claude-code-usage-go-up-heres-the-measurement-edk <p>I kept hitting Claude Code's weekly limit, so before keeping any token-saving tool I now measure it on a fixed task set.</p> <h2> How I measured </h2> <ul> <li>5 fixed tasks: fix a bug, scan a log, check git changes, summarize a CSV, summarize test results</li> <li>3 runs per task, same day, same test repo, <code>claude -p --output-format json</code> </li> <li>Cost-weighted tokens = input 1, cache write 1.25, cache read 0.1, output 5</li> <li>Answer quality counted as part of correctness</li> </ul> <h2> Results </h2> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Tool</th> <th>What it does</th> <th>Cost-weighted tokens vs baseline</th> <th>Correct</th> </tr> </thead> <tbody> <tr> <td>rtk v0.51.0</td> <td>compresses shell output via a hook</td> <td>+7.1%</td> <td>15/15</td> </tr> <tr> <td>caveman v3.1.0</td> <td>makes replies ultra-terse</td> <td>+10.9%</td> <td>15/15</td> </tr> <tr> <td>pxpipe v0.14.0</td> <td>renders bulky context as images via a local proxy</td> <td>+196.7%</td> <td>15/15</td> </tr> </tbody> </table></div> <h2> Why all three went up </h2> <ol> <li> <strong>Fewer tokens per step can mean more steps.</strong> rtk trimmed output by ~63%, but Claude ran extra commands to see what was cut. Each turn re-reads the context, so cache reads rose 19%.</li> <li> <strong>Anything added to the system prompt is paid on every call.</strong> caveman's rules (~2,900 tokens) cost more than the 12% output it saved.</li> <li> <strong>Images are not free.</strong> pxpipe's rendered context cost more than the original text; cache writes on the first call went from ~27K to ~102K.</li> </ol> <h2> Where my tokens actually go </h2> <p>About 42% of my cost is the always-loaded context (system prompt, instructions, tool list). Command output is about 6%.</p> <p>Raw numbers and method: <a href="https://github.com/awesomefred0827-hash/claude-token-log" rel="noopener noreferrer">https://github.com/awesomefred0827-hash/claude-token-log</a></p> <h2> Next: vendor and patch </h2> <p>I'm thinking of vendoring all three (pulling the source and patching it myself):</p> <ul> <li> <strong>rtk</strong>: keep error lines uncompressed</li> <li> <strong>caveman</strong>: one "be brief" rule, no skill list</li> <li> <strong>pxpipe</strong>: compress only past a context-size threshold</li> </ul> <p>What do you think? Would these patches work, or would you fix them differently? I'll re-test whatever you suggest and post the results.</p> claude ai productivity discuss