← ClaudeAtlas

agent-toolsmithlisted

Cut the token bill of AI coding agents by turning code they keep rewriting into tested tools - audit session logs and scratch scripts for repeated commands and scripts (scripts/find_repeats.py), rank them by sessions x tokens, build the top ones as one command each with --dry-run, tests and a clear stop on failure, list them in an AGENT_TOOLS.md catalog, add a «tools first» rule for the agents, and re-audit monthly. Use when agents keep writing the same scripts, when a project burns tokens on routine (merges, restarts, DB queries, screenshots, reports), or when asked to make reusable tools for agents.
roama-studio/skills · ★ 0 · AI & Automation · score 72
Install: claude install-skill roama-studio/skills
# Agent toolsmith A script the agent runs costs only its output in context. A script the agent writes again costs the whole code, the retries and the debugging - every session. Repeats are the cheapest tokens to save. ## 1. Audit (read-only) ``` python scripts/find_repeats.py --transcripts ~/.claude/projects/<project-folder> \ [--scripts <scratch folders> <agent run folders>] --days 30 --top 20 ``` It reports shell-command shapes used in 3+ sessions (paths, numbers and strings masked, plain reads and output filters dropped), scripts with identical code in 2+ sessions, script names reused in 3+ sessions, each with uses and rough tokens. Then look at the top rows yourself: - **what** repeats (one line), **how often** (sessions = the main measure), **who** writes it, **what it costs** (tokens, minutes, the same mistakes again - encoding errors, forgotten steps); - does a tool **already exist** that the agent did not find? Then the fix is the catalog, not new code; - one-off edit scripts (`fix_*`, `patch_*`) are not candidates - they are how big edits are done. Write the audit as a table and pick the top 3-5 with the owner of the project before building. ## 2. Build each tool - One command, one job, named by the job: `merge_branch <branch>`, `restart_backend`, `db peek "select ..."`. - `--dry-run` that prints every step it would take; read-only where possible. - Stops at the first failure and prints what to do next (the exact manual command), never half-finishes s