Master Developer Prompt Kit

buildagentic · McDis Framework

One prompt. One Claude call. Full engineering package from any transcript.

Update: Copilot coding agent now uses fewer premium requests — each session consumes just one premium request.

---

Files

File	Purpose
master_dev_prompt.txt	The system prompt — defines schema + Claude's role
run_master_dev.sh	Shell script to pipe any transcript through Claude Code or Codex (supports --mock or MOCK_OUTPUT=1 to emit outputs/sample.json when a model CLI is unavailable)
validate_output.py	JSON validator for structure + enum checks
master_dev_runtime.py	Shared Python runtime for prompt loading and strict output validation
batch_run_master_dev.sh	Batch processor for all transcripts/*.txt files
Makefile	Local/CI shortcuts (validate-file, validate-outputs, ci)
.github/workflows/validate-master-dev-prompt.yml	GitHub Actions validation workflow
sample_transcript.txt	Test transcript (Dear Saigon SMS ordering system)
app.py	FastAPI HTTP server — POST /process calls Claude directly via SDK, GET / serves the web UI
static/index.html	Terminal-style web UI — paste a transcript, get rendered artifacts

---

Usage

1. Make the script executable (first time only)

chmod +x run_master_dev.sh

2. Run with your transcript

./run_master_dev.sh your_transcript.txt

3. Save output as JSON

./run_master_dev.sh your_transcript.txt > output.json

4. Test with the sample

./run_master_dev.sh sample_transcript.txt

The script auto-detects:

• claude CLI first
• falls back to codex exec if claude is not installed

Optional reliability env vars:

• MAX_RETRIES (default 3)
• RETRY_DELAY_SECONDS (default 2)
• MASTER_DEV_TIMEOUT_SECONDS (default 0, disabled)

Example:

MAX_RETRIES=5 RETRY_DELAY_SECONDS=3 MASTER_DEV_TIMEOUT_SECONDS=90 \
./run_master_dev.sh sample_transcript.txt > output.json

5. Validate output JSON

python3 validate_output.py output.json

Validation is strict:

• required object keys must match exactly
• actions.items, implementation_plan.milestones, implementation_plan.tech_tasks, and code_suggestions.snippets must each contain at least one item
• enum fields must use the exact allowed values

Or with Make:

make validate-file FILE=output.json

6. Batch run all transcripts

Put .txt transcripts in transcripts/, then run:

chmod +x batch_run_master_dev.sh
./batch_run_master_dev.sh

Custom folders:

./batch_run_master_dev.sh ./my_transcripts ./my_outputs

Validate all generated outputs:

make validate-outputs

Batch behavior:

• successful runs write outputs/<name>.json
• schema failures write outputs/<name>.invalid.json
• run logs go to outputs/<name>.log
• make validate-outputs validates only *.json and ignores *.invalid.json

7. CI checks

Local CI-equivalent check:

make ci

GitHub Actions runs the same checks on push and pull request.

Python test suite:

python3 -m pytest -q

---

Manual (no script)

{
  cat master_dev_prompt.txt
  cat your_transcript.txt
  echo ""
  echo '"""'
} | claude

Codex alternative:

{
  cat master_dev_prompt.txt
  cat your_transcript.txt
  echo ""
  echo '"""'
} | codex exec --skip-git-repo-check -

---

Output Modes

Mode	What you get
design_doc	context, requirements, architecture, alternatives, decisions, open questions
pm_summary	plain-language overview, scope, timeline implications
actions	tasks with owner, priority (low/medium/high), type
implementation_plan	milestones with ETA + risks, tech tasks with complexity (S/M/L)
code_suggestions	real runnable snippets in detected language, stack context

---

FastAPI Server

Run the included server for HTTP access and a browser UI:

export ANTHROPIC_API_KEY=sk-...
pip install fastapi uvicorn anthropic python-dotenv
uvicorn app:app --reload

• GET / — opens the web UI (paste transcript, get rendered artifacts)
• POST /process — returns schema-validated structured JSON from any transcript
• POST /process/stream — streams partial text and emits a final schema-validated result or validation error event
• GET /health — liveness check

Optional: set SERVER_API_KEY to require an X-Api-Key header on all requests.

Minimal SDK snippet

import anthropic, json

client = anthropic.Anthropic(api_key=YOUR_KEY)

with open("master_dev_prompt.txt") as f:
    system_prompt = f.read()

def run_master_dev(transcript: str) -> dict:
    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=8096,
        system=system_prompt,
        messages=[{"role": "user", "content": f'{transcript}\n"""'}]
    )
    return json.loads(response.content[0].text)

---

Autopilot Watcher

Run the watcher to automatically process any transcript dropped into transcripts/:

python3 watcher.py

Then drop any .txt file into transcripts/ — validated JSON output appears in outputs/ automatically.

Custom directories:

python3 watcher.py --transcripts ./my-transcripts --outputs ./my-outputs

Stop with Ctrl+C. Files already processed (with a matching .json in outputs/) are skipped automatically. Schema failures are written to outputs/<name>.invalid.json, and validator details are appended to outputs/<name>.log.

---

GitHub Actions Workflows

Three workflows run automatically on the main branch:

Workflow	File	Trigger	Purpose
Validate Master Dev Prompt	validate-master-dev-prompt.yml	Push / PR to main	Installs Python deps, runs make ci (schema validation + pytest suite)
Copilot Auto-Fix	copilot-autofix.yml	Push / PR	GitHub Copilot automated code fix suggestions
Cleanup Old Workflow Runs	cleanup-workflow-runs.yml	Weekly (Sun 3am UTC) + manual	Deletes workflow run history older than 30 days via GitHub API

Running CI locally

make ci

Triggering the cleanup workflow manually

1. Go to Actions → Cleanup Old Workflow Runs
2. Click Run workflow → select branch main → Run workflow

Branch conventions

• Default branch: main
• Copilot feature branches are ephemeral — they are deleted after their PR is merged or closed
• Do not force-push to main