Reversing the Development Ratio: AI‑First Planning with Compound Engineering

Core Idea: Inverted Development Ratio

Traditional software development spends most time on mechanical coding (≈80%) and little on planning, leading to inefficiency and inconsistent code. The new workflow inverts this ratio, making planning the primary activity and delegating execution to AI agents.

Key Plugin – Compound Engineering

By installing the /plugin marketplace add EveryInc/compound-engineering-plugin developers can use simple markdown files (e.g., plan.md) and voice commands to drive the entire development process without a traditional IDE.

Primary Commands and Execution Logic

/ce:plan

– Starts a deep planning phase where AI generates a structured plan, including problem definition, solution outline, file list, and checklist‑style acceptance criteria. /ce:brainstorm – Launches multiple research agents (code‑base, documentation, external best‑practice) that read the repository, retrieve historical solutions, and aggregate findings into plan.md. /ce:work – Parses plan.md, then automatically performs coding, testing, and verification in sequence. If context is lost, a new session can attach to the same plan.md to resume from a checkpoint.

The workflow achieves roughly 80% of time spent on AI planning and 20% on agent execution, keeping all core decisions in the planning file.

Voice Input – Hands‑Free Interaction

With large‑language‑model (LLM) agents, voice transcription errors are corrected by contextual understanding, allowing developers to issue commands like /ce:plan or describe requirements verbally. Recommended tools:

Monologue (EveryInc) – streams live voice to Claude Code.

WhisperFlow – high‑accuracy transcription for development commands.

Using a professional shotgun microphone, developers can dictate planning instructions, requirement descriptions, and more, dramatically reducing manual typing.

Parallel Multi‑Session Execution

Up to 4–6 independent Ghostty windows can run separate Claude Code sessions simultaneously:

One session generates new plans.

Another executes code based on an existing plan.

A third conducts research on recent industry trends.

A fourth handles bug‑fixes.

Switching between /ce:plan and /ce:work across sessions maximizes time usage while keeping each session isolated; GitHub version control safeguards any code changes.

Essential Configuration Tweaks

Claude Code’s default permission prompts must be adjusted for smooth operation. Add the following to ~/.claude/settings.json:

{
  "permissions": {
    "allow": ["WebSearch","WebFetch","Bash","Read","Write","Edit","Glob","Grep","Task","TodoWrite"],
    "deny": [],
    "defaultMode": "bypassPermissions"
  },
  "skipDangerousModePermissionPrompt": true
}

Enable a task‑completion sound hook:

{
  "hooks": {
    "Stop": [{
      "hooks": [{
        "type": "command",
        "command": "afplay /System/Library/Sounds/Blow.aiff"
      }]
    }]
  }
}

Configure Zed editor auto‑save (every 500 ms) to let Claude Code monitor file changes in real time:

{
  "autosave": {
    "after_delay": {
      "milliseconds": 500
    }
  }
}

Pre‑Research with Real‑Time Information

Before planning, run /last30days (provided by the open‑source Granola tool) to fetch the latest 30‑day content from Reddit, X, YouTube, Hacker News, etc. This yields up‑to‑date industry data, e.g., comparing token usage of Vercel agent‑browser (82‑93% reduction) versus Playwright (13 700 tokens).

Structuring Meeting Content

Transcribed meeting or interview recordings can be fed into Claude Code via /ce:plan, automatically producing a structured plan.md that includes goals, user stories, technical solutions, and milestones. The Granola‑based MCP protocol enables seamless conversion without manual copy‑paste.

All‑Scenario Applicability

The workflow extends beyond code generation to strategic documents, product specifications, and competitive analysis. The unified loop is communication input → plan generation → iterative optimization , suitable for any task requiring structured output.

Remote Execution on a Mac Mini

Deploy OpenClaw and Claude Code on a Mac Mini, integrate with Telegram, and issue commands from a mobile device, e.g.:

Send /ce:plan fix the timeout problem via Telegram.

The device creates a plan.md and syncs it to the Zed editor.

Receive the plan via AgentMail for offline preparation.

Conclusion

By centering planning and leveraging AI agents for execution, the Compound Engineering workflow restructures development, reduces manual coding, and improves consistency. When consistently applied, it forms a standardized, traceable development system that benefits both individual developers and teams in the AI era.