Prompt Builder Skill

You are the prompt-builder skill. Your role is to build complete agent prompts by calling prompt_builder.py, which handles everything deterministically.

Overview

This skill builds complete agent prompts by calling a Python script that:

• Reads specializations from database (task_groups.specializations)
• Reads context from database (context_packages, error_patterns, reasoning)
• Reads full agent definition files from filesystem
• Applies token budgets per model
• Validates required markers are present
• Saves prompt to file and returns JSON result

Prerequisites

• Database must be initialized (bazinga/bazinga.db exists)
• Config must be seeded (run config-seeder skill first at session start)
• Agent files must exist in agents/ directory

When to Invoke This Skill

• RIGHT BEFORE spawning any BAZINGA agent
• When orchestrator needs a complete prompt for Developer, QA Expert, Tech Lead, PM, Investigator, or Requirements Engineer
• Called ON-DEMAND to get the latest context from database

Your Task

When invoked, you must:

Step 1: Read Parameters File

The orchestrator writes a params JSON file before invoking this skill. Look for it at:

bazinga/prompts/{session_id}/params_{agent_type}_{group_id}.json

Example: bazinga/prompts/bazinga_20251217_120000/params_developer_CALC.json

Params file format:

{
  "agent_type": "developer",
  "session_id": "bazinga_20251217_120000",
  "group_id": "CALC",
  "task_title": "Implement calculator",
  "task_requirements": "Create add/subtract functions",
  "branch": "main",
  "mode": "simple",
  "testing_mode": "full",
  "model": "haiku",
  "output_file": "bazinga/prompts/bazinga_20251217_120000/developer_CALC.md"
}

Additional fields for retries:

{
  "qa_feedback": "Tests failed: test_add expected 4, got 5",
  "tl_feedback": "Error handling needs improvement"
}

Additional fields for CRP (Compact Return Protocol):

{
  "prior_handoff_file": "bazinga/artifacts/bazinga_20251217_120000/CALC/handoff_developer.json"
}

Additional fields for PM spawns:

{
  "pm_state": "{...json...}",
  "resume_context": "Resuming after developer completion"
}

Step 2: Call the Python Script

Run the prompt builder with the params file:

python3 .claude/skills/prompt-builder/scripts/prompt_builder.py --params-file "bazinga/prompts/{session_id}/params_{agent_type}_{group_id}.json"

The script will:

1. Read all parameters from the JSON file
2. Build the complete prompt
3. Save prompt to output_file path
4. Output JSON result to stdout

Step 3: Return JSON Result to Orchestrator

The script outputs JSON to stdout:

Success response:

{
  "success": true,
  "prompt_file": "bazinga/prompts/bazinga_20251217_120000/developer_CALC.md",
  "tokens_estimate": 10728,
  "lines": 1406,
  "markers_ok": true,
  "missing_markers": [],
  "error": null
}

Error response:

{
  "success": false,
  "prompt_file": null,
  "tokens_estimate": 0,
  "lines": 0,
  "markers_ok": false,
  "missing_markers": ["READY_FOR_QA"],
  "error": "Prompt validation failed - missing required markers"
}

Return this JSON to the orchestrator so it can:

1. Verify success is true
2. Read prompt from prompt_file for the Task spawn
3. Check markers_ok is true

Step 4: IMMEDIATELY Spawn Agent (CRITICAL - SAME TURN)

🔴 DO NOT STOP after receiving JSON. IMMEDIATELY call Task() to spawn the agent.

After verifying success: true, spawn the agent in the SAME assistant turn:

Task(
  subagent_type: "general-purpose",
  model: "{haiku|sonnet|opus}",
  description: "{agent_type} working on {group_id}",
  prompt: "FIRST: Read {prompt_file} which contains your complete instructions.
THEN: Execute ALL instructions in that file.
Do NOT proceed without reading the file first."
)

🚫 ANTI-PATTERN:

❌ WRONG: "Prompt built successfully. JSON result: {...}" [STOPS - turn ends]
   → Agent never spawns. Workflow hangs until user says "continue".

✅ CORRECT: "Prompt built successfully." [IMMEDIATELY calls Task() with prompt_file]
   → Agent spawns automatically. Workflow continues.

The entire sequence (params file → prompt-builder → Task spawn) MUST complete in ONE assistant turn.

Params File Reference

prior_handoff_file Behavior:

• If path is valid and file exists: Handoff section added with instruction to read file
• If path is invalid (traversal attempt, wrong pattern): Warning logged, section omitted
• If path is valid but file doesn't exist: Warning logged, section omitted (agent proceeds without prior context)
• Path validation: Must start with bazinga/artifacts/, match handoff_*.json pattern, no path traversal (../)

What the Script Does Internally

1. Reads parameters from JSON file
2. Queries database for task_groups.specializations → reads template files
3. Queries database for context_packages, error_patterns, agent_reasoning
4. Reads full agent definition file (agents/*.md) - 800-2500 lines
5. Applies token budgets per model (haiku=900, sonnet=1800, opus=2400)
6. Validates required markers are present (e.g., "READY_FOR_QA", "NO DELEGATION")
7. Saves prompt to output_file
8. Returns JSON result to stdout

Error Handling

If the result has success: false, do NOT proceed with agent spawn. Report the error to orchestrator.

Legacy CLI Mode (Backward Compatibility)

The script still supports direct CLI invocation for manual testing:

python3 .claude/skills/prompt-builder/scripts/prompt_builder.py \
  --agent-type developer \
  --session-id "bazinga_123" \
  --branch "main" \
  --mode "simple" \
  --testing-mode "full"

Add --json-output to get JSON response in CLI mode.