Does "rtk" skill cut agent tokens by 60–90%? We tested it

I'll draft five analytical paragraphs from only the facts you provided—no invented versions, dates, or figures.JetBrains published a test of the rtk skill aimed at Claude Code agent workflows, asking whether it can cut agent tokens by 60–90%. The write-up is on the JetBrains AI blog under a July 2026 path and was also listed on Hacker News under HN AI Agents, where it sat at 3 points with 0 comments at the time of the summary.

The post frames rtk as a skill used inside Claude Code rather than as a separate product surface. The central claim under test is a large reduction in agent tokens—stated in the title as a 60–90% cut—so the technical stake is whether skill-level behavior can shrink the volume of tokens an agent burns while still doing useful work. No separate architecture diagram, benchmark suite name, or product version is given in the available facts; the signal is the token-savings claim tied to Claude Code usage.

For engineers and builders running agent loops, token volume is often the main cost and latency lever. A skill that reliably trims agent tokens by tens of percent would change how often teams can afford long tool-using sessions, how they size budgets per task, and how aggressively they chain multi-step agent runs. Even if the upper end of the range does not hold in every setup, a measured drop in that band is large enough to matter for anyone already watching Claude Code spend closely.

Market context is still thin: the HN thread had only 3 points and no comments when summarized, so there is little public pushback or corroboration yet. The source being JetBrains’ AI blog, not a third-party lab, means readers should treat the 60–90% figure as a vendor-tested claim about rtk on Claude Code until independent runs appear. Interest in agent cost control is already high; a skill-shaped fix is a different pitch from model swaps or prompt-only compression.

Practical takeaway: treat the 60–90% range as a hypothesis to re-measure on your own Claude Code tasks with and without rtk, using the same workloads and logging total agent tokens. Watch for follow-up discussion on the HN item and for whether JetBrains publishes the exact method behind the savings so others can reproduce it. Until then, the useful next step is a controlled A/B on a real agent job, not adopting the headline number as a default planning assumption.JetBrains published a test of the rtk skill aimed at Claude Code agent workflows, asking whether it can cut agent tokens by 60–90%. The write-up is on the JetBrains AI blog under a July 2026 path and was also listed on Hacker News under HN AI Agents, where it sat at 3 points with 0 comments at the time of the summary.

The post frames rtk as a skill used inside Claude Code rather than as a separate product surface. The central claim under test is a large reduction in agent tokens—stated in the title as a 60–90% cut—so the technical stake is whether skill-level behavior can shrink the volume of tokens an agent burns while still doing useful work. No separate architecture diagram, benchmark suite name, or product version is given in the available facts; the signal is the token-savings claim tied to Claude Code usage.

For engineers and builders running agent loops, token volume is often the main cost and latency lever. A skill that reliably trims agent tokens by tens of percent would change how often teams can afford long tool-using sessions, how they size budgets per task, and how aggressively they chain multi-step agent runs. Even if the upper end of the range does not hold in every setup, a measured drop in that band is large enough to matter for anyone already watching Claude Code spend closely.

Market context is still thin: the HN thread had only 3 points and no comments when summarized, so there is little public pushback or corroboration yet. The source being JetBrains’ AI blog, not a third-party lab, means readers should treat the 60–90% figure as a vendor-tested claim about rtk on Claude Code until independent runs appear. Interest in agent cost control is already high; a skill-shaped fix is a different pitch from model swaps or prompt-only compression.

Practical takeaway: treat the 60–90% range as a hypothesis to re-measure on your own Claude Code tasks with and without rtk, using the same workloads and logging total agent tokens. Watch for follow-up discussion on the HN item and for whether JetBrains publishes the exact method behind the savings so others can reproduce it. Until then, the useful next step is a controlled A/B on a real agent job, not adopting the headline number as a default planning assumption.