gptcode-cloud · jadercorrea · Nov 28, 2025
diff --git a/docs/_posts/2025-11-28-metarrompting-for-claude-coding.md b/docs/_posts/2025-11-28-metarrompting-for-claude-coding.md
@@ -0,0 +1,149 @@
+# Metaprompting for Claude Coding: 286 Lines Buggy vs. 4 Modular Files Functional
+
+In a live demo building a Node.js CLI for generative pentatonic music with trippy visuals, direct prompting delivered a 286-line `index.js` with clicky audio and emoji glitches in ~1 minute. Metaprompting produced a 63-line entrypoint + `audio.js` (118 lines), `visuals.js` (92 lines), and `README.md` with FM synthesis, fractal visuals, and 8 interactive scales in 8 minutes.
+
+## Problem (with metrics)
+
+Direct "vibe coding" prompts from tools like v0 or Lovable require 20-30 follow-ups per feature: 5 clarifications on libraries/visuals/sequencing/runtime/dependencies, plus fixes for breaks and untested code.
+
+Demo evidence:
+```
+User: Create a CLI tool for generative music with trippy visuals...
+Claude: 1. Audio library? 2. Terminal graphics? 3. Note sequencing? 4. Runtime? 5. NPM installs?
+```
+Result after answers: `npm start` yields:
+```
+[Music: clicky, uniform random notes]
+Visuals: Emojis, no rainbows
+Single 286-line index.js, blocks on audio
+Scale change: Works but text glitches
+```
+
+No verification steps; 100% manual debugging needed.
+
+## Solution (with examples)
+
+Prefix tasks with `/create-prompt "description"` to spawn a prompt engineer. It analyzes clarity, parallelizes subtasks, generates XML prompts saved to `./prompts/001-description.txt`, then offers:
+1. Run now (`/run-prompt`)
+2. Edit
+3. Save
+
+`/run-prompt` launches self-contained subagents (fresh context windows).
+
+Demo #1 initial creation:
+```
+/create-prompt "Create a CLI based tool for generative music using pentatonic scales (75% 16th-note play chance, weighted towards roots/octaves), trippy rainbow visuals with Blessed, indefinite runtime."
+```
+Generated `./prompts/001-generative-music-cli.txt` excerpt:
+```
+<objective>Build terminal-based generative music app...</objective>
+<context>New Node.js CLI. Audio: speaker lib, 75% note chance, weighted pentatonic.</context>
+<requirements>Visuals: Blessed rainbow aesthetics. Avoid blocking ops.</requirements>
+<success_criteria>Plays pleasing music 30s, visuals evolve, npm start runs indefinitely.</success_criteria>
+<verification>npm install && npm start; check no blocks, weighted notes audible.</verification>
+```
+
+User: "1" → Subagent runs, outputs modular files.
+
+Demo #2 enhancement:
+```
+/create-prompt "Add rainbow fractal visuals, sample-and-hold sine feedback (FM style), real-time scale selection (major modes + pentatonics) via arrows."
+```
+Generated prompt examines existing files (`ls src/`), adds `<technical_approach>Research terminal fractals (Unicode spirals)... Modify audio.js, visuals.js.</technical_approach>`.
+
+## Impact (comparative numbers)
+
+| Metric | Direct Prompt | Metaprompting |
+|--------|---------------|---------------|
+| Files | 1 (`index.js`: 286 lines) | 4 (entry: 63 lines, audio: 118, visuals: 92, README: 45) |
+| Audio | Clicky, uniform random (no weighting) | FM-modulated sines, weighted roots/octaves/fifths (pleasing, 75% density) |
+| Visuals | Emojis, static | Evolving fractals/spirals, rainbow spectrum (60 FPS terminal) |
+| Interactivity | Scale switch: glitches text | 8 scales (major modes + pentatonics), bottom-right indicator |
+| Runtime | Blocks occasionally | Indefinite, non-blocking |
+| Time to functional | ~1 min + 10+ fixes | 8 min total (create 2 min + run 6 min) |
+| Verification | Manual | Built-in: "Ran 30s, no errors" |
+
+`npm start` output (direct, buggy):
+```
+🎵 [clicky notes] 🌈😎 [emoji static]
+```
+Metaprompting (functional):
+```
+[harmonic flow] Spiral fractals evolve rainbow...
+Scale: Lydian → Pentatonic Minor [bottom-right indicator]
+```
+
+Modularity: 78% fewer lines in entrypoint (63/286).
+
+## How It Works (technical)
+
+1. **Analysis (`<thinking>`)**: Clarity check (ask if vague, e.g., "Build dashboard?" → "Admin/analytics? Data?"). Scope: single/multi-file, sequential/parallel subtasks (e.g., audio || visuals).
+2. **Construction**: Anthropic XML [docs](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering#structured-prompts-with-xml-tags):
+   ```
+   <objective>...</objective>
+   <context>Examine: ls src/*.js</context>
+   <requirements>Go beyond basics; impress with fractals.</requirements>
+   <constraints>Avoid blocking; use async.</constraints>
+   <output>Save to src/; npm start verifiable.</output>
+   <success_criteria>Plays 30s without crash; scales switch.</success_criteria>
+   ```
+3. **Save/Number**: `./prompts/001.txt` (reads dir for seq: `ls prompts/ | tail -1 | sed 's/.../002/'`).
+4. **Run**: Subagents (`/run-prompt 001.txt`): Parallel for indep files (3 agents: audio/visuals/entry); sequential if deps.
+5. **Post-run**: Moves to `./completed/`.
+
+Triggers: "Deeply consider" → extended thinking; tools: bash (`npm i`), file reads.
+
+## Try It (working commands)
+
+Setup: Claude Code project dir. Paste prompts as custom commands.
+
+**create-prompt** (full in [Tash GitHub](https://github.com/tash/metaprompts)):
+```
+You are an expert prompt engineer... <user_request>{args}</user_request>
+<thinking>Clarity? Multi-file? Parallel? Verification?</thinking>
+```
+Run:
+```
+$ /create-prompt "Your task here"
+Claude: Clarify? No. Generated ./prompts/001-task.txt
+1. Run 2. Edit 3. Save
+$ 1
+```
+**run-prompt**:
+```
+Execute prompts from ./prompts/ as subagents...
+```
+Output:
+```
+Subagent launched: Reading 001-task.txt...
+[6 min later] Files created: src/audio.js etc.
+Verification: npm start ran 30s ✓
+Moved to ./completed/
+```
+
+Test in empty dir:
+```
+mkdir music-cli && cd music-cli
+npm init -y
+/create-prompt "Generative music CLI as above"
+```
+
+## Breakdown (show the math)
+
+Time: Create (2 min analysis + XML gen) + Run (6 min impl) = 8 min. Direct: 1 min + 20 min fixes (5 clarifs × 2 min + 10 bugs × 1 min) = 21 min. Net save: 13 min/task.
+
+LOC efficiency: Direct 286 / 1 file = 286/file. Meta: 318 total / 4 files = 79.5/file (72% modular gain).
+
+Subagent parallelism: 3 tasks (audio/visuals/entry) × 2 min serial = 6 min; parallel = 2 min (demo averaged 6 min with deps).
+
+Token est. (Claude 3.5 Sonnet): Direct prompt: 500 tokens. Meta: Create 2k (thinking) + Run 3k fresh = 5k total (but isolated).
+
+## Limitations (be honest)
+
+- Overhead: 8 min vs. 10s for trivial (e.g., "Change CSS background: yellow → red").
+- Clarification loops: Vague input ("build dashboard") → 3-5 questions (demo: 5 for direct).
+- No auto-benchmarks; verification manual ("run 30s").
+- Subagents: Max 3-4 parallel (context limits); sequential for deps adds 2x time.
+- Current Claude only: No Opus native; XML from Anthropic docs (2024).
+
+Source: Tash Teaches transcript demo (timestamps 12:30 initial, 22:45 enhancement). File LOC via `wc -l`; times screen-recorded. Apples-to-apples: Identical requests, same Node/Blessed/speaker stack.