Low-attention workflow for experiments, internal tools, prototypes, and mechanical migrations. Define a machine-checkable finish line, then loop autonomously until it is met. Use when work is off the critical path and the model's judgment can stand in for the human's.
Installs into .claude/skills of the current project.
Sponsors
SSkills DirectorySkills Directory
Know which skills are safe — weekly.
SSkills DirectorySkills Directory
Is that skill safe to install?
Are you the author of Oneshot?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/nickolasbenakis-oneshot)
---
name: oneshot
description: Low-attention workflow for experiments, internal tools, prototypes, and mechanical migrations. Define a machine-checkable finish line, then loop autonomously until it is met. Use when work is off the critical path and the model's judgment can stand in for the human's.
---
# Oneshot
Spend model intelligence instead of human attention. The human should be able
to start this and forget it until the result is ready.
## Steps
1. **Finish line first.** Write down, before any code, a check a script can
evaluate: a test suite passes, a CLI command prints the expected output,
CI is green, a parity check between old and new implementation passes.
If no such check exists, the first task is building it.
2. **Scaffold** in a new branch or directory `oneshot/<slug>`, isolated from
critical-path code.
3. **Loop**: implement → run the finish-line check → read failures → fix.
Don't ask the human questions; make a reasonable choice and note it in
`DECISIONS.md` in the working directory.
4. **Large mechanical work** (ports, migrations, codemods): build a small
harness instead of brute-forcing it:
- enumerate units of work in dependency order (e.g. walk the import graph)
- hand each unit to a subagent with a brief written with the `delegate`
skill
- a check that fails if the new side falls behind the old (counts,
parity tests)
5. **Review before reporting**: low attention means the human won't catch what
you miss. Once the finish line passes, have the `playbook-reviewer`
subagent (or another fresh subagent) review the diff
for correctness bugs, especially on real-world inputs the fixtures don't
cover (other quote styles, redirects, encodings). Fix confirmed findings.
6. **Stop conditions**: finish line met and reviewed → report. Same failure 3
times, or the approach is clearly wrong → stop and report what you learned
and what you'd try next. Never loop indefinitely.
## Tools
- Claude Code: `/loop` (self-paced) or a workflow script for fan-out.
- bb: a child thread in a worktree; or `bb automation create --in 1m ...` to
run it off the main thread entirely.
## Done
One message: what was built, where it is, the finish-line result, and the
contents of `DECISIONS.md`.