Conforming Workspace Walkthrough

To see how PROCESS.md operates in practice, let’s look at the sample workspace provided in this repository under /examples/company_ops/.

This workspace models a marketing and growth ops department automating experiment analysis and marketing readiness reviews.

Directory Layout

The workspace follows this structured setup:

examples/company_ops/
├── pdt.yaml                          # Global project config
├── processes/                        # Executable workflows
│   ├── growth_experiment_review/
│   │   └── PROCESS.md
│   └── marketing_launch_readiness/
│       └── PROCESS.md
├── skills/                           # Reusable agent capabilities
│   ├── experiment-analysis/
│   │   └── SKILL.md
│   └── positioning-review/
│       └── SKILL.md
├── tools/                            # Executable integrations
│   ├── experiment_lookup/
│   │   ├── tool.yaml
│   │   └── main.py
│   └── campaign_asset_lookup/
│       ├── tool.yaml
│       └── index.js
└── schemas/                          # Data format contracts
    └── experiment-review.schema.json

Workspace Settings: `pdt.yaml`

The project root defines environment rules and tool gates:

project:
  id: company_ops
  name: Company Operations Workflows

paths:
  processes: ./processes
  skills: ./skills
  tools: ./tools
  schemas: ./schemas

tools:
  allow:
    - experiment_lookup
    - campaign_asset_lookup

policy:
  dry_run_blocks_side_effects: true
  require_approval_for_external_writes: true

Step-by-Step Breakdown: `growth_experiment_review`

Let’s review the growth_experiment_review process (processes/growth_experiment_review/PROCESS.md).

Part 1: Frontmatter & Description

The process defines its identity, owner, and its operational boundaries:

---
id: growth_experiment_review
name: Growth Experiment Review and Analysis
version: 0.1.0
owner: growth-ops
status: active
description: Analyzes completed product growth experiments and updates downstream spreadsheets.
tags: [growth, analytics, operations]
runtime: pdt.process.v0
---

# Description

This process governs the review of completed product growth experiments.
It ensures that experiment metadata is cross-checked with telemetry, and that
outcomes are summarized, formatted, and validated before publication.

## Scope
- Covers all user-facing product experiments on web and mobile.
- Excludes architectural backend experiments.

Part 2: Step-by-Step Workflow Execution

Step 1: Ingest Experiment Metadata

## Step 1: Retrieve Experiment Metadata
Look up the experiment metadata using the `tool/experiment_lookup` tool. 
Specify the experiment identifier provided in the workspace context.
If no experiment matches, halt execution immediately.

Runtime Behavior: The compiler binds the experiment_lookup tool to this step. The LLM is prohibited from performing general analysis or making other API calls; its single job is to call the lookup tool.

Step 2: Perform Data Analysis

## Step 2: Analyze Results
Evaluate the telemetry data using the guidelines in `skill/experiment-analysis`.
Format the final output payload representing the analysis summary.

Runtime Behavior: The compiler parses SKILL.md for experiment-analysis and appends its guidelines to the step prompt. No active tools are provided here, reducing reasoning drift.

Step 3: Conformance Validation

## Step 3: Validate and Format Output
Ensure the final output conforms strictly to the schema structure defined in `schema/experiment-review`.
Verify that key metrics (e.g. lift, p-value, sample size) are populated.

Runtime Behavior: The runtime intercepts the output of this step and validates the JSON payload against experiment-review.schema.json. If it fails, the step is re-run or flagged with an error.

The Reusable Skill: `experiment-analysis`

The file skills/experiment-analysis/SKILL.md provides detailed instructions for calculating statistical metrics:

---
id: experiment-analysis
name: Product Growth Experiment Analysis
version: 1.0.0
---

# Guidelines

When analyzing product growth experiments, follow these mathematical constraints:
1. **Sample Size Check:** Ensure the control group and variation group each contain at least 1,000 unique users.
2. **P-Value Threshold:** The result is statistically significant only if the p-value is `< 0.05`.
3. **Lift Calculation:** Calculate lift as `(Variation Mean - Control Mean) / Control Mean * 100`.

The Executable Tool: `experiment_lookup`

The folder tools/experiment_lookup/ exposes a programmatic Python script to the agent:

`tool.yaml`

id: experiment_lookup
name: Lookup Experiment Metrics
description: Fetches raw experiment metrics and telemetry data from the database.
runtime: python3
entrypoint: main.py
parameters:
  type: object
  properties:
    experiment_id:
      type: string
      description: The unique identifier of the experiment (e.g. EXP_2026_01).
  required:
    - experiment_id
side_effects: false

Because side_effects is false, this tool can run without triggering human approval blocks.