Open Steps makes coding agents report in plain language, with 441 stars

Pavlo Kharmanskyi has published Open Steps, an MIT-licensed pack of eight Claude Code skills that force the coding agent to report in plain language, ask its questions in words a non-engineer can answer, and end each piece of work with a one-screen verdict. The repository was created on 25 August 2026, has 441 stars and 75 forks, and was last pushed on 14 September.

What the pack does

The eight skills each run at a specific point in a session. os-done-or-not writes a one-screen report at the end of the work with a verdict: done or not, anything needed from the person, any new work left for later, safe to close. os-step-by-step produces numbered steps a non-technical person can follow, and only after the agent has tried everything itself. os-ask-simple states a question in plain words, names what it costs later, and marks one recommendation. os-check-work refuses to trust another session's report and checks every claim against what actually happened. os-say-simple rewrites any text without losing facts or bad news. os-big-picture keeps one permanent file that names every feature, how far each got, and what is queued.

The author, who describes himself as not an engineer, says he built the pack because his own coding agent kept reporting its work in engineering shorthand, so he could not tell whether it was done.

Numbers the author quotes

The activation numbers were remeasured on 12 September against a set of 25 phrases, three runs each, headless, on three Claude models. Kharmanskyi reports that all 25 phrases triggered the right skill 100 percent of the time on Opus 5, 98 percent on Sonnet 5, and 85 percent on Haiku 4.5. Off-topic questions triggered zero of nine on every model. He also states the pack costs 770 tokens per session for the always-on skill descriptions, a figure Claude Code reports itself with claude plugin details open-steps.

These are the author's own numbers on his own test set, and he says so in the README. He also points out that Haiku results change between runs, that three runs per phrase is a quick check rather than a benchmark, and that on Haiku two skills are unreliable enough that a user will sometimes have to type the skill name by hand.

The pack is a Claude Code plugin, and the skills also install into Codex, Cursor and Gemini CLI through their shared skills folder. Two hooks, both plain shell scripts, keep the skills running across a long session.

Anyone running AI-assisted development for a team that mixes engineers and non-engineers has the exact class of problem this pack targets. The engineer reads the diff. The product owner reads a report they cannot verify. Open Steps rewrites that report as a plain verdict so the product owner can supervise the work directly. It also runs another agent against the first one's claims in os-check-work, which is the pattern a team building agents on top of agents already needs. Whether the 85 to 100 percent activation numbers stay true on a different phrase set is the first thing anyone adopting this should measure for themselves, and the evaluation harness is in evals/ for that purpose.

Source

Primary source: kharmanskyi/open-steps on GitHub.