Software Engineer
6 items tagged with this
The full playbook for cheap, sharp agent work. Get observability in first (statusline, burndown, OTEL), compress what reaches the model (rtk, headroom), route who does the work (premium delegates, cheap executes, advisor escalation, model-per-task), then work the long tail: caveman, the right browser validator, scoped scans, fewer MCPs, handover, and a memory system. The capstone of the Token Saving series.
Browser-harness is a coordinate/CDP-driven browser harness for AI agents. Preconfigure the site navigation once and the agent reuses the structure instead of regenerating scripts every run, which keeps token usage down on repeated browser work.
Oracle is a CLI and Claude Code skill for getting a second opinion from a stronger model. A lesser model drives, escalates the hard calls to a premium one, and adversarial review goes through the same door. Pay premium only for the 10% that needs it.
Ponytail is a Claude Code plugin that forces the laziest solution that works. YAGNI, stdlib first, one line over fifty. Less code generated means fewer output tokens, smaller diffs, and less to read back later.
Two token trimmers that stack: rtk is a Rust CLI proxy that filters dev-command output at the shell (60-90% off ls, grep, git, tests), and Headroom is a local compression proxy that compresses tool output, logs, files and RAG chunks on the wire, with CacheAligner to protect your prompt cache. Real numbers from two days of use.
How to stop getting rate-limited after an hour on Claude Code, Codex, or Copilot. Context window mechanics, the /effort command, model routing, kicking new conversations, and the silent token drains most people miss.