nebi-promptinglisted
Install: claude install-skill romanized/nebi-skill
# Prompt sharpening: preserve the voice
The active harness's direct-instruction override always wins.
## Choose the requested outcome
An explicit prompt, sharpen, split-into-prompts, or review-before-execution request uses the artifact and gate below. If this skill loads for an ordinary action request, organize the instructions internally and carry out authorized work; rough wording alone does not create a `Go?` gate. Preserve any actual ambiguity or approval boundary.
## Bound the context
Choose the local-primer mode before reading anything:
- **General sharpening:** read no local state. References such as `our skill`, `our project`, or `our two Nebi skills` remain General unless paired with explicit prior-work language; they do not authorize any local read. Words such as `attached`, `supplied`, `provided`, `local`, `working directory`, or `current directory` describe inputs the future executor receives and also remain General on their own.
- **Ongoing local work:** activated only by an explicit reference to prior joint work on this machine, such as “this repo” or “we're already working in”; a bug or prior-progress mention alone stays General. When active, do both allowed checks before drafting when available: read only the selected task checkpoint identified by the startup context or explicitly supplied by the user, and if the current directory is a git repo, run `git status --short --branch`. If no checkpoint is selected, use the user's words; do not search other tas