| name | code-analysis-playbook |
| description | How to use the code execution sandbox well — parse the JSON payload (not disk files), compute a couple of useful diagnostics before forecasting, and keep the session stateful within a turn. Load this before writing code. No scripts. |
Code-analysis playbook
A short guide to using the run_code sandbox productively. This is a starter
skill — extend it with the diagnostics that matter for your problem.
Where your data lives
All data comes from the JSON payload in your context — there are no disk
files and no network. The history arrives as a CSV string (e.g.
target_history_csv). Parse it with io.StringIO, never as a file path:
import io, pandas as pd
df = pd.read_csv(io.StringIO(payload["target_history_csv"]))
The sandbox is stateful within a turn: parse once in your first code block,
then reuse the DataFrame in later blocks instead of re-parsing.
Compute before you forecast
Run a couple of cheap diagnostics so your forecast is grounded in arithmetic,
not vibes:
- Recent trend — slope/return over the last N observations.
- Volatility — recent standard deviation of changes; it sets how wide your
quantile bands should be.
- Sanity check — does your point forecast sit within a plausible multiple
of recent moves? If not, revisit it.
Use the printed numbers to set the point forecast and to calibrate the spread
between your low and high quantiles — wider when recent volatility is high.
Domain focus (edit this for your use case)
For monthly food CPI, the series is an index (base 2002=100) that moves slowly
and seasonally; month-over-month changes are typically small. Use recent
year-over-year change and seasonality, not a fixed guess, to set your trajectory
and the width of your interval bands.
Room to grow
- Add your own diagnostic patterns (regime detection, seasonality, covariates).
- Drop reusable reference values into a
references/ file and load_skill_resource them.
