runwaylisted
Install: claude install-skill ruijiezh67/ruijie_skills
# Runway
Quota is fuel. Never take off on a leg you cannot land: a task that hits the
limit halfway costs its tokens twice, once to do it and once to rebuild the
context afterwards.
## 1. Read the gauges (always first)
Run `python <skill-dir>/scripts/pace.py`. It prints a tier and both windows:
used %, time to reset, burn rate in %/hour, hours until cap.
- `UNKNOWN` or "no samples": the status line recorder is not set up. Give the
user the setup from `references/setup.md` and plan from `/usage` numbers they
paste. Do not guess quota.
- A value printed `N/A` is unmeasured. Say so. Never treat it as 0.
## 2. Size every task
For each task in the backlog, write one row: task, kind, size, needs-model.
- **kind**: `model` (Claude does the work), `wall` (runs without Claude: a
training job, Colab, CI, a sweep, a reviewer replying), or `mixed`.
- **size** in % of a 5-hour window: S under 5, M 5 to 20, L over 20. Estimate
from the measured burn rate times expected active hours. If there is no
burn rate yet, say the size is a guess.
- **interruptible**: can it stop halfway with a usable checkpoint?
## 3. Order by these rules
Apply in order. The ones above win.
1. **Start an L task only if headroom covers it with a 30% margin**
(`100 - used ≥ 1.3 × size`). Otherwise schedule it as the first thing
after the next reset.
2. **Launch `wall` jobs before the cap, not after.** At AMBER or RED, the most
valuable thing left is starting work that runs while Claude is