Meridian Model Building Skill

This skill guides the user through the process of creating a Meridian model, accumulating the code into a Python script.

Core Workflow

Interactivity Checkpoint Rule

Throughout this workflow, you will encounter CRITICAL INTERACTIVE CHECKPOINTs. At each checkpoint, you MUST:

1. Present the current proposed configurations, parameters, mappings, script path, or status to the user for approval.
2. Ask the user if they are ready to proceed using the available user-interaction tool (e.g., ask_question), structured as a multiple-choice question. Do NOT use raw chat text.
3. Wait for their response before proceeding.
   • MANDATORY: You MUST pause at every checkpoint regardless of the initial prompt instructions (even if the user request contains phrases like "run autonomously", "execute directly", "fix autonomously", etc.). The initial request does NOT bypass these interactive checkpoints.
   • Note: If the user replies to a checkpoint with a generic approval (e.g., "proceed", "do what you think is best"), proceed with the proposed defaults.

Python Environment and Meridian Source

Before running any Python, pick the interpreter of the environment where google-meridian is installed, and use it for every command and script below:

1. If the user named an environment, use it. Otherwise, if a virtual environment is already active (VIRTUAL_ENV or CONDA_PREFIX is set), use python3. Otherwise, if the project directory (the current working directory) has a virtual environment (.venv/ or venv/), use its interpreter (e.g., .venv/bin/python3). Otherwise use python3.
2. Check it with <interpreter> -c "import meridian". If that fails, ask the user which environment to use instead of searching the machine or installing packages unprompted.

Steps below ask you to read Meridian modules (e.g., meridian.model.spec) to confirm argument names and defaults. Users usually install Meridian from PyPI (pip install google-meridian), so its source is not in their project folder. Find the installed package and read the files there, or print signatures and docstrings with inspect:

<interpreter> -c "import meridian, os; print(os.path.dirname(meridian.__file__))"

1. Initial Setup

• Prompt the user for the input CSV file path, the desired path for the generated Python script, the EDA HTML report output path, and the saved model path (meridian_model.binpb by default). If the user does not specify output paths, default to model_build/ in the active project directory (or relative to the input data directory) for the script and all outputs (meridian_model.binpb, eda.html).
• CRITICAL INTERACTIVE CHECKPOINT: Present the gathered paths to the user and obtain confirmation before proceeding to data loading.

2. Add Data Loading & Column Mapping Code

• Target Module: meridian.data.data_frame_input_data_builder
• Action:
  • Check CSV Format: Before loading data, verify if the CSV data is in the right format. Use the meridian-doc-consultant skill to find the data-loading guide on developers.google.com/meridian that matches the columns observed in the data (e.g., the "load geo-level data without reach and frequency" or "load geo-level data with organic and non-media" guide) to understand the expected columns and data types. Consult references/csv_format_reference.md for details on expected row/column structure and data quality guardrails. If the format is incorrect or missing required columns, attempt to autonomously convert the dataset to the expected format for the user (e.g., renaming columns, restructuring) unless you are uncertain and need user input.
  • Read the header row of the provided CSV using Python to get the column names.
  • Propose heuristic mappings based on column keywords (e.g., 'sales' -> kpi_col, 'spend' -> media_spend_cols) and infer the kpi_type ('revenue' or 'non_revenue') based on the columns (e.g., 'revenue' or 'sales' implying 'revenue', and 'conversions' or 'leads' implying 'non_revenue').
  • Robust Mapping: If the user prompt specifies mapping a column name that does not exist in the CSV, do not assume it is a literal name if it looks like a description (e.g., 'media_impressions' vs 'ChannelX_impression'). Use heuristics to find matching columns and proceed.
  • Present the proposed mapping to the user.
  • CRITICAL INTERACTIVE CHECKPOINT: Present the proposed column mappings to the user and obtain approval before continuing to model configuration.
  • Accumulate the data loading code using meridian.data.data_frame_input_data_builder.DataFrameInputDataBuilder and its with_* methods (e.g. with_kpi, with_media). See data_builder_template.md.

3. Add Model Configuration Code

• Target Modules: meridian.model.spec, meridian.model.model
• Action:
  • Read the ModelSpec and PriorDistribution definitions in meridian.model.spec.
  • Guide the user through configuration, prompting for relevant values while explaining their purpose based on the source code docstrings.
  • CRITICAL INTERACTIVE CHECKPOINT: Present the proposed model specification parameters to the user and obtain approval before continuing.
  • Accumulate the code to initialize meridian.model.spec.ModelSpec and meridian.model.model.Meridian. See model_spec_template.md.
  • Accumulate code: mmm.sample_prior()

4. Add Exploratory Data Analysis (EDA) Code

• Target Module: meridian.model.eda.meridian_eda
• Action:
  • Read meridian_eda.py or module docstrings to confirm the generate_and_save_report method.
  • Accumulate code to initialize meridian_eda.MeridianEDA and call generate_and_save_report(filepath) using the user's specified path.
  • CRITICAL INTERACTIVE CHECKPOINT: Present the EDA output path configuration and obtain approval before proceeding to the model fitting step.

5. Add Model Fitting Code

• Target Module: meridian.model.model
• Action:
  • Read the sample_posterior method in meridian.model.model to understand its parameters.
  • Prompt the user for MCMC parameters: n_chains, n_adapt, n_burnin, n_keep.
  • CRITICAL INTERACTIVE CHECKPOINT: Present the MCMC parameters to the user and obtain approval before proceeding to compile the model fitting code.
  • Accumulate code: mmm.sample_posterior(...)

6. Add Model Saving Code

• Target Module: meridian.schema.serde.meridian_serde
• Action:
  • Generate code to save the model using meridian_serde.save_meridian() to the user-specified path (or the default). See script_template.md.
  • Default Filename: The default filename for the saved model is meridian_model.binpb (in the model_build/ directory). Use this filename if the user does not specify a model filename, even if the script file is named differently.
  • Skip Sampling Handling: If the user requests to skip fitting or posterior sampling, still include the model saving step (meridian_serde.save_meridian(mmm, save_path)) using the initialized Meridian model object so the output model file is always created.
  • WARNING: Do NOT use the deprecated meridian.model.model.save_mmm function. Use meridian_serde.save_meridian exclusively.
  • CRITICAL INTERACTIVE CHECKPOINT: Present the model save path and filename to the user and obtain approval before proceeding to script execution.

7. Execution & Script Setup

• Action:
  • Write the accumulated Python script to the user-specified path. If your file-writing tool can ask the user to review the written file, don't request that review: the next checkpoint already asks for approval, and an extra prompt pauses the workflow.
  • CRITICAL INTERACTIVE CHECKPOINT: Ask the user for final confirmation to execute the model building script now.
  • Artifact Preservation: When completing a task that requires generating outputs (like scripts, models, or reports), do NOT delete these generated artifacts at the end of your turn. They are the deliverables requested by the user. Only clean up truly temporary scratch files if necessary.
    • Path Handling for Outputs: In generated scripts, use the output paths confirmed in Step 1, relative to the project directory the script is run from (or absolute paths).
  • Execute the Script:
    • Always run the script with the project directory as the working directory (not a subdirectory), so relative paths resolve as confirmed.
    • Use the interpreter chosen in "Python Environment and Meridian Source" above.
    • Example command: .venv/bin/python3 model_build/my_model.py
  • Handling Long Runs: If the command is sent to the background due to execution time, wait for the background task to complete and check the final output to catch runtime errors.
  • Differentiated Error Handling:
    • If it's a Syntax Error or Import Error, read the relevant source code to understand the correct usage or interface.
    • If it's a ValueError or parameter constraint violation (e.g., knots too large), check the docstring of the class/function or consult the meridian-doc-consultant skill to find valid values in the documentation.
    • Autonomy: If a fix requires changing configuration, prompt the user for confirmation. If the user response grants autonomy, proceed to fix it.

8. Conclusion

• Action:
  • Confirm execution success and artifact generation.