Skip to content

Instantly share code, notes, and snippets.

@jmanhype
Last active December 23, 2025 19:55
Show Gist options
  • Select an option

  • Save jmanhype/a9a3670bf7dfba774681790099363829 to your computer and use it in GitHub Desktop.

Select an option

Save jmanhype/a9a3670bf7dfba774681790099363829 to your computer and use it in GitHub Desktop.
AI Coding Assistants & Frameworks: A Complete Guide - Claude Code, Codex, Beads, Speckit, Quint Code, and more

AI Coding Assistants: What I Actually Use

TL;DR for Pranav: Claude Code is my primary driver. Codex is good but works best as a second opinion, not the main tool. Here's my complete setup.


The Short Answer

What do I prefer? Claude Code, hands down.

Why not Codex alone? Codex is solid but:

  • It shines at grinding through well-defined specs
  • It's great as a reviewer catching things Claude missed
  • But for reasoning, debugging, and complex multi-file work — Claude Code wins

My actual workflow:

1. Claude Code implements → does the heavy lifting
2. Codex reviews → catches edge cases I missed
3. Repeat until both agree it's solid

They complement each other. Don't pick one — use both in tandem.


Getting Started (If You're New)

Step 1: Install Claude Code

npm install -g @anthropic-ai/claude-code

This is your primary tool. Use it for everything first.

Step 2: Use Codex for Review

After Claude Code finishes a feature, run Codex on the same code:

codex "Review this code for edge cases and bugs"

Step 3: Add Beads for Memory

cargo install beads

This gives your agents persistent task memory across sessions. Game changer for long projects.

Step 4: Get Promptlet (macOS)

Download from josh.ing/promptlet Quick-insert prompt templates into any AI tool with a hotkey.


Why Claude Code > Codex (for primary work)

Aspect Claude Code Codex
Reasoning Excellent — traces through complex logic Good but shallower
Context Deep codebase awareness More isolated
Debugging Best-in-class Decent
Autonomy Interactive, collaborative Better for "set and forget"
Long specs Good Great — can grind for hours

Bottom line: Claude Code for thinking, Codex for grinding.


My Full Stack

Here's everything I use, organized by purpose:

Primary Agents

  • Claude Code — main driver for implementation
  • Codex CLI — second opinion, code review, long-running tasks

Productivity Boosters

  • Promptlet — macOS app for quick prompt template insertion
    • Hotkey → search → insert pre-crafted prompts instantly
    • Includes: Ultrathink, Chain of Thought, Step-by-Step, Deep Analysis, SOLID Principles
    • Works with Claude, ChatGPT, Gemini, any text field

Memory & Task Tracking

  • Beads — persistent git-backed task memory for agents

Specification Workflow

  • Speckit — spec-driven development with Beads + Pivotal Labs TDD

Reasoning Frameworks

Prompting Techniques

  • Verbalized Sampling — 2-3x diversity improvement in LLM outputs
    • Training-free prompting strategy to mitigate mode collapse
    • Model-agnostic (GPT, Claude, Gemini, Llama)
    • Great for creative writing, synthetic data, dialogue simulation

Quality & Evaluation

  • Gentleman Guardian Angel — AI pre-commit hook
  • Bloom — automated behavior evaluation for LLMs
    • 4-stage pipeline: Understand → Ideate → Rollout → Judge
    • Test for sycophancy, bias, oversight-subversion, etc.
    • Multi-model comparison (OpenAI, Anthropic, Bedrock)
    • Interactive web viewer for transcript analysis

Agent Teams

Environment Setup


For Your Data Science Work (Box 3)

Since you're working on causal inference and PyReason integration, these are especially relevant:

  • AI Data Science Team — agents for EDA, feature engineering, ML pipelines
  • Modeltime — time series forecasting in R (10K+ series/day)
  • Beads — track your experiments across sessions
  • Verbalized Sampling — generate diverse synthetic training data
  • Bloom — evaluate model behavior systematically

Quick Start Recommendation

  1. Today: Install Claude Code + Promptlet
  2. This week: Try Claude Code as your primary. Use Codex only for review.
  3. Next week: Add Beads for task persistence.
  4. When ready: Add Speckit for structured spec → plan → implement workflow.

You'll feel the difference immediately. Claude Code thinks with you. Codex executes for you.


All Resources

Tool Link Purpose
Claude Code npm i -g @anthropic-ai/claude-code Primary coding agent
Codex CLI OpenAI Review, autonomous tasks
Promptlet josh.ing/promptlet Quick prompt templates (macOS)
Beads steveyegge/beads Persistent agent memory
Speckit jmanhype/speckit Spec-driven workflow
Spec-Kit github/spec-kit GitHub's official version
Quint Code m0n0x41d/quint-code Reasoning framework
Deep Truth Mode QuantumCousin/Deep-truth-mode-spirit First principles
Verbalized Sampling CHATS-lab/verbalized-sampling Output diversity
Bloom safety-research/bloom Behavior evaluation
GGA Gentleman-Programming/gentleman-guardian-angel Pre-commit AI review
Oh-My-OpenCode code-yeongyu/oh-my-opencode Agent orchestration
AI Data Science Team business-science/ai-data-science-team Data science agents
Modeltime business-science/modeltime Time series forecasting
ACFS Dicklesworthstone/agentic_coding_flywheel_setup VPS bootstrap

Hit me up if you have questions — happy to pair on setting any of this up.

— Jay, December 2025

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment