Stress-test your development plan
before launching agents.

You are helping describe a software project for the M4 calculator, a phase-based model of one person working with several AI agents.

First ask clarifying questions if any of these are undefined: a shared acceptance criterion for the manual and AI-assisted scenarios, task dependencies, baseline effort without AI, how AI will be used, and mandatory human checks. Keep uncertainty explicit and do not present an estimate as a forecast.

For external calibration, use only empirical studies published in 2026; prefer June–August 2026 when available. Do not use coefficients from older experiments as defaults. If a recent source measures commits, merge rate, or time to merge rather than end-to-end time to an accepted result, mark the estimate as a proxy and explain the conversion.

Then:
1. Decompose the project into verifiable tasks with independent outcomes wherever that is substantively possible.
2. Do not add dependencies merely to impose an order: predecessors must reflect a real inability to start the task earlier.
3. Group similar tasks into classes. For each class, estimate aiCycleMultiplier = end-to-end time to an accepted result with AI / manual baseline. A value of 0.7 means a cycle 30% shorter than manual work; 1.2 means 20% longer.
4. Estimate humanFraction = the share of active human time inside the AI cycle, not a share of the manual baseline. For example, with k=0.7 and humanFraction=0.4, the human is active for 0.28 of the manual baseline and the agent for 0.42. Do not derive humanFraction automatically from k: it is a separate assumption.
5. Add evidence for every class and set kind to measured for directly comparable timing measurements, proxy for indirect metrics such as throughput, or assumption for expert judgment. Add publishedAt in YYYY-MM format for an external source; omit it for an unpublished assumption. If no reliable 2026 source exists, explicitly use assumption. Briefly explain the human share in humanFractionBasis.
6. Parameters c and gamma cannot be below 1. If coordination overhead is unknown, use 1 and explicitly warn that this is optimistic.
7. Return exactly one JSON object without Markdown or comments. IDs must be unique ASCII identifiers. Every predecessor must reference an existing ID, and the graph must be acyclic.

Result schema:
{
  "schemaVersion": "m4-calculator/v1",
  "projectName": "Project name",
  "timeUnit": "hours",
  "p": 3,
  "c": 1.0,
  "gamma": 1.0,
  "taskClasses": [
    {
      "id": "feature",
      "name": "Convenient class name",
      "aiCycleMultiplier": 0.7,
      "humanFraction": 0.4,
      "evidence": {
        "kind": "assumption",
        "label": "Expert estimate for this project",
        "note": "Why this coefficient was selected"
      },
      "humanFractionBasis": "Human work that remains inside the cycle"
    }
  ],
  "tasks": [
    {
      "id": "task-1",
      "classId": "feature",
      "name": "Optional short name",
      "description": "Optional description and acceptance criterion",
      "baselineEffort": 16,
      "predecessors": []
    }
  ]
}