A Claude Code skill that teaches the agent to emit
ARF records as it works. Reasoning lands
in .arf/ (an orphan git branch) keyed to commit SHAs. Reviewers see
what the agent thought, not just what it changed.
# 1. Install the arf CLI (this is what the skill shells out to).
cargo install arf-cli
# 2. Install the skill via skillz.
skillz install github:ducks/arf-skillOr, without skillz: drop SKILL.md into Claude Code's skills directory
manually.
After install, whenever Claude is making non-trivial changes it runs
arf record with structured what/why/how/backup fields. Records persist
to .arf/ and stay correlated with the commits they led to.
To inspect a project's reasoning trail:
arf log # recent records
arf graph # git history interleaved with reasoning
arf diff <commit> # single commit with reasoning + diff
arf browse # TUI browserAI agents modify code, make decisions, take actions. Today their reasoning is buried in chat logs or lost entirely. ARF captures it in a structured form that:
- Humans can review in PRs and postmortems
- Other agents can consume in multi-agent workflows
- Tools can validate and process
The skill is the integration that makes ARF emission automatic instead of a thing you have to remember to invoke.