Python Model Code Generator

Purpose

Generate Python modeling code from a validated candidate method pool and cleaned data.

This skill implements the approved candidate methods (baseline, main candidates, optional improved) as Python scripts. It should produce code that reads cleaned data, runs the planned models, saves results in the experiments/roundN/ output structure, generates a run_summary.json, and produces portable artifacts for downstream experiment reports, robustness checks, and paper writing.

This skill does not choose the model, clean raw data, review code, run final QA, or write paper sections.

When to use

Use this skill:

• After model-code-analyzer hands off to python-model-code-generator.
• When implementation.target = python.
• When Python is allowed by the contest or requested by the user.
• When the candidate method pool and cleaned data are ready.
• When executable .py scripts are needed for one or more candidate methods in a specific experiment round.

Preconditions

The following should already exist or be provided:

• A validated problem parse.
• A validated problem classification artifact.
• A candidate method pool at methods/Qx/qx_method_candidates.md.
• implementation.target = python.
• Cleaned data under workspace/data/data_clean/, unless the model does not need tabular data.
• Data audit report and field mapping if data is used.
• code/model-code-analyzer.md — the code thinking document specifying round number, methods to implement, and output structure.
• Expected result files and figure files from the candidate method pool.
• Runtime notes if specific dependency or portability constraints exist.

If the candidate method pool is missing, hand back to method-selector.

If code/model-code-analyzer.md is missing, hand back to model-code-analyzer.

If cleaned data is required but missing, hand back to data-auditor-cleaner.

If implementation.target is not python, hand back to model-code-analyzer.

Inputs

Use or request:

• methods/Qx/qx_method_candidates.md — candidate method pool for each subquestion.
• code/model-code-analyzer.md — code thinking document.
• workspace/data/data_report.md, if available.
• Cleaned data under workspace/data/data_clean/.
• Field mapping from the data audit stage.
• Required model inputs per candidate method.
• Expected result outputs per candidate method.
• Expected figure outputs per candidate method.
• The round number (N for roundN).
• Runtime notes such as:
• minimal-dependencies
• portable-artifacts
• avoid-notebook-only-code
• save-csv-json-png
• fixed-random-seed

Workflow

1. Read the code thinking document and candidate method pool.

• Identify the target subquestion(s) and round number.
• Identify which methods are to be implemented in this round (baseline, main candidates, optional improved).
• Identify required inputs, expected outputs, validation needs, and robustness needs per method.
• Do not change the selected methods or modeling route.

1. Confirm Python target.

• Confirm implementation.target = python.
• Preserve all runtime notes.
• Use a minimal, common scientific Python stack (numpy, pandas, scipy, scikit-learn, matplotlib).

1. Check data readiness.

• Use cleaned data from workspace/data/data_clean/.
• Do not read from workspace/data/data_raw/ unless only inspecting metadata is explicitly requested.
• Do not overwrite raw data.
• Confirm required fields match the data audit report.

1. Plan script structure per subquestion.

• Save Python code under code/Qx/ (e.g., code/Q1/, code/Q2/).
• One script per candidate method: qx_m1_baseline.py, qx_m2_main.py, qx_m3_improved.py.
• A README.md in each code folder: run order, inputs, outputs, dependencies.
• A run_all.py only if batch execution is useful.
• Prefer plain .py scripts over notebook-only workflows.
• Keep scripts runnable from the project root.
• Use small helper functions when they improve clarity.

1. Generate code for each candidate method.

• Input: Read from workspace/data/data_clean/.
• Processing: Implement the mathematical logic from the candidate method pool.
• Save results: Save to results/Qx/experiments/roundN/tables/.
• Save figures: Save to results/Qx/experiments/roundN/figures/.
• Save metrics: Save to results/Qx/experiments/roundN/metrics/.
• Write log: Write execution log to results/Qx/experiments/roundN/logs/.
• Write run_summary.json: At end of run_all.py or the last script, generate results/Qx/experiments/roundN/run_summary.json.

1. Implement baseline code first.

• Implement the baseline method first.
• Save baseline outputs separately (e.g., m1_scores.csv).
• Make baseline outputs comparable with other methods.

1. Implement main and optional methods.

• Implement each candidate method with the same output conventions.
• Use variable names that match the candidate method specification where practical.
• Save intermediate values needed for explanation or robustness checks.

1. Implement run_summary.json generation.

• After all methods have executed, write run_summary.json with:

```python import json from datetime import datetime

runsummary = { "question": "Q1", "round": "round1", "executiontimestamp": datetime.now().isoformat(), "implementationtarget": "python", "randomseed": SEED, "methods": [ { "methodid": "M1", "methodname": "equalweightscoring", "script": "code/Q1/q1m1baseline.py", "status": "success", "executiontimeseconds": 12.5, "inputfiles": ["workspace/data/dataclean/indicatordata.csv"], "outputfiles": ["results/Q1/experiments/round1/tables/m1scores.csv"], "figurefiles": ["results/Q1/experiments/round1/figures/m1scorebar.png"], "metricssummary": {"scorerange": [0.45, 0.92], "rankingtop3": ["A", "B", "C"]}, "errors": [], "warnings": [] } ], "datainputs": ["workspace/data/dataclean/indicatordata.csv"], "environment": { "pythonversion": "3.10", "keydependencies": ["numpy", "pandas", "matplotlib", "scipy"] } } with open('results/Q1/experiments/round1/runsummary.json', 'w') as f: json.dump(runsummary, f, indent=2, ensure_ascii=False) ```

1. Add a README.md per code folder.

• State the subquestion and round.
• List scripts and their roles.
• State run order.
• List expected input data files.
• List expected output paths.
• State required dependencies.
• State known portability risks.

1. Hand off to code-reviewer.

• The router should route .py scripts to python-code-reviewer.

Outputs

Produce Python implementation artifacts such as:

• code/Q1/README.md
• code/Q1/q1_m1_baseline.py
• code/Q1/q1_m2_main.py
• code/Q1/q1_m3_improved.py (if applicable)
• code/Q1/run_all.py (if useful)
• results/Q1/experiments/roundN/ (created by running the scripts):
• tables/*.csv
• figures/*.png
• metrics/*.json
• logs/*.log
• run_summary.json

Output format

Prefer this JSON-compatible summary:

{
  "python_code_generation_summary": {
    "implementation_target": "python",
    "subquestion": "Q1",
    "round": "round1",
    "runtime_notes": ["minimal-dependencies", "portable-artifacts"],
    "generated_scripts": [
      "code/Q1/q1_m1_baseline.py",
      "code/Q1/q1_m2_main.py",
      "code/Q1/run_all.py"
    ],
    "data_inputs": [
      "workspace/data/data_clean/indicator_data.csv"
    ],
    "result_outputs": [
      "results/Q1/experiments/round1/tables/m1_scores.csv",
      "results/Q1/experiments/round1/tables/m2_scores.csv",
      "results/Q1/experiments/round1/metrics/m1_metrics.json",
      "results/Q1/experiments/round1/metrics/m2_metrics.json"
    ],
    "figure_outputs": [
      "results/Q1/experiments/round1/figures/m1_score_bar.png",
      "results/Q1/experiments/round1/figures/m2_weight_bar.png"
    ],
    "run_instructions": [
      "cd code/Q1 && python run_all.py"
    ],
    "dependencies": ["numpy", "pandas", "matplotlib", "scipy"],
    "known_risks": [
      "Column names must match the data audit report before interpretation."
    ],
    "recommended_next_skill": "code-reviewer"
  }
}

Python coding rules

Use clear, minimal, reviewable Python code.

Prefer:

• plain .py scripts
• pathlib.Path
• numpy, pandas, scipy, scikit-learn (when justified), matplotlib
• relative paths from project root
• explicit result saving to results/Qx/experiments/roundN/
• fixed random seeds (SEED = 2026)
• portable outputs: .csv, .json, .png
• run_summary.json generation

Avoid unless explicitly approved:

• notebook-only code
• hard-coded local absolute paths
• hidden data cleaning inside modeling scripts
• heavy deep learning frameworks (torch, tensorflow)
• unnecessary third-party packages
• Python-only output formats as the only result artifact
• code that only prints results without saving them
• complex abstractions that make contest debugging harder

Common accepted dependencies:

numpy
pandas
scipy
scikit-learn
matplotlib

Avoid heavy dependencies unless explicitly justified:

torch
tensorflow
xgboost
lightgbm
large geospatial packages
specialized optimization packages

Path rules

Use relative paths where practical.

Recommended paths:

workspace/data/data_clean/
code/Qx/
results/Qx/experiments/roundN/
  ├── tables/
  ├── figures/
  ├── metrics/
  ├── logs/
  └── run_summary.json

Avoid hard-coded absolute paths.

Example path setup:

from pathlib import Path

ROOT = Path(".")
DATA_DIR = ROOT / "workspace" / "data" / "data_clean"
RESULT_DIR = ROOT / "results" / "Q1" / "experiments" / "round1"
TABLE_DIR = RESULT_DIR / "tables"
FIGURE_DIR = RESULT_DIR / "figures"
METRIC_DIR = RESULT_DIR / "metrics"
LOG_DIR = RESULT_DIR / "logs"

for d in [TABLE_DIR, FIGURE_DIR, METRIC_DIR, LOG_DIR]:
    d.mkdir(parents=True, exist_ok=True)

Artifact rules

• Read cleaned data from workspace/data/data_clean/.
• Do not modify files under workspace/data/data_raw/.
• Save Python scripts under code/Qx/.
• Save numeric outputs under results/Qx/experiments/roundN/tables/.
• Save figures under results/Qx/experiments/roundN/figures/.
• Save metrics under results/Qx/experiments/roundN/metrics/.
• Save logs under results/Qx/experiments/roundN/logs/.
• Generate run_summary.json under results/Qx/experiments/roundN/.
• Use clear file names with method prefix:
• m1_scores.csv, m2_weights.csv, m1_metrics.json
• m1_score_bar.png, m2_weight_distribution.png
• Separate outputs from different methods clearly.
• Save important intermediate results if later paper explanation or robustness checks require them.

Randomness rules

If randomness is used:

• Set fixed seeds: random.seed(SEED), np.random.seed(SEED).
• Record the seed in comments and run_summary.json.
• For scikit-learn: pass random_state=SEED where supported.

import random
import numpy as np

SEED = 2026
random.seed(SEED)
np.random.seed(SEED)

Data handling rules

• Use pandas for tabular data when practical.
• Keep column names aligned with the data audit report.
• Do not silently drop columns or rows.
• Do not silently impute missing values unless the method plan or data report allows it.
• Keep feature-target separation explicit for prediction tasks.
• Avoid train-test leakage.
• Keep units and indicator directions consistent with the method plan.

Plotting rules

• Use matplotlib by default.
• Save figures with plt.savefig(FIGURE_DIR / "m1_result.png", dpi=300, bbox_inches='tight').
• Use svg.fonttype = 'none' for editable SVG text.
• Use readable labels, units, legends, and titles.
• Close figures after saving: plt.close(fig).
• Do not rely on interactive display.

import matplotlib.pyplot as plt

plt.rcParams['svg.fonttype'] = 'none'
fig, ax = plt.subplots()
ax.plot(x, y)
ax.set_xlabel("x")
ax.set_ylabel("y")
fig.tight_layout()
fig.savefig(FIGURE_DIR / "m1_result.png", dpi=300, bbox_inches='tight')
fig.savefig(FIGURE_DIR / "m1_result.svg", bbox_inches='tight')
plt.close(fig)

Commenting rules

Use comments to explain:

• data inputs and their source paths
• key variables and their meanings
• model steps and non-obvious formulas
• output files and their destinations
• assumptions inherited from the method plan
• random seed and reproducibility notes

Avoid over-commenting trivial syntax.

Rules

• Do not select or change the model — implement what the candidate method pool specifies.
• Do not clean raw data.
• Do not fabricate data, labels, parameters, metrics, or results.
• Do not write paper text.
• Do not perform final QA.
• Do not use notebook-only workflows.
• Do not use Python-only binary artifacts as the only saved result.
• Do not hide assumptions inside code.
• Do not silently drop rows or columns.
• Do not overwrite raw data.
• Always generate run_summary.json.
• Always create the full experiments/roundN/ directory structure.
• Do not claim the code is reviewed or correct before python-code-reviewer.
• Separate baseline outputs from main method outputs.

Verification

Before handing off, verify:

• implementation.target = python.
• Every generated script maps to a method in the candidate method pool.
• Baseline code exists for every subquestion unless explicitly impossible.
• Scripts are saved under code/Qx/.
• Scripts write outputs to results/Qx/experiments/roundN/.
• Cleaned data paths are used.
• No raw data file is overwritten.
• Random seeds are fixed where needed.
• run_summary.json generation is implemented.
• Figures are saved to results/Qx/experiments/roundN/figures/.
• Outputs are portable enough for downstream checks.
• A README.md exists in the code folder.
• Dependency risks are listed.
• The next skill is code-reviewer.

Failure modes

Stop and report a blocker if:

• The candidate method pool is missing.
• code/model-code-analyzer.md is missing.
• implementation.target is not python.
• Cleaned data is required but unavailable.
• Required data fields are missing.
• The selected method cannot be implemented with reasonable Python dependencies.
• The expected output is not defined.
• Generating code would require inventing data, labels, parameters, or metrics.
• Contest rules disallow Python for official computation and no mixed-language workflow is approved.

Stop conditions

This skill must stop instead of guessing when:

• Model variables cannot be mapped to cleaned data fields.
• Important parameters are unspecified and cannot be safely defaulted.
• The method requires a heavy dependency that is not approved.
• The code would need to overwrite raw data.
• A script would produce results that cannot be traced to the candidate method pool.
• Continuing would change the mathematical meaning of the model.
• Python is not allowed by contest rules for the intended stage.

When stopping, output:

• blocker
• why it matters
• affected subquestion or model
• missing information or artifact
• safe partial implementation if any
• recommended next skill
• next action

Handoff

After generating Python scripts, hand off to:

code-reviewer

The handoff should include:

• generated .py script paths
• code folder README.md path
• candidate method pool path
• cleaned data paths
• field mapping
• expected result paths under results/Qx/experiments/roundN/
• expected figure paths
• runtime notes
• dependency notes
• run instructions
• known risks and assumptions

code-reviewer should route the scripts to `python-co

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: zhnnky329
• Source: zhnnky329/MathModeling-skills
• License: MIT

Install and usage instructions live in the source repository linked above.