packlightlisted
Install: claude install-skill Syedadil0803/packlight
# packlight
Most tokens in a long session are the chat being re-read on every step. Short sessions that start from
a small, current set of docs cost far less. Load only the file the command needs. (`/packlight new-project`, `weigh`, `unpack` with a space mean the same.)
| Command | Do this |
|---|---|
| `/packlight` | No code in the folder: read `new-project.md`. Code: read `setup.md`. |
| `/packlight-new-project` | Read `new-project.md`, even if a few files exist. |
| `/packlight-weigh` | Run `python3 <this skill folder>/scripts/usage_report.py` from the project folder (no other file needed). No docs/STATUS.md yet: run it without `--since`; that is the "before" baseline. Otherwise add `--since YYYY-MM-DD` for a before/after split: the date the user gives, else the `--since` date setup wrote in STATUS.md, else the first git commit of docs/STATUS.md, else ask the user (file dates are unreliable). Show the numbers and the script's tips. It never prints message content. |
| `/packlight-unpack` | Read `unpack.md`. |
## Model per task
Use the cheapest model that does the job well. When this session runs on a stronger model than a step
needs, hand that step to a subagent on the cheaper model and keep only its summary.
| Work | Model tier (current example) |
|---|---|
| Stack choice, architecture, security, hard bugs | Strongest (Opus) |
| Interview, writing CLAUDE.md and feature pages, feature work, reviews | Mid (Sonnet) |
| Reading and searching code, drift checks, line count