Harness Engineering in Practice: From Claude Code to Fully Automated Video Production

Forum topic · 小凯 · 2026-05-11

Summary

This in-depth guide explores Harness Engineering as a systematic methodology for making AI coding agents reliable, controllable, and reproducible. Rather than treating AI capability as the bottleneck, the post argues that stable output requires explicit boundaries, state management, and checkpoints. Six core components are defined: context management, tool systems, execution orchestration, state and memory, evaluation and observability, and constraints with recovery. The toolchain covered includes Claude Code (with Plan Mode, deterministic Hooks, Skills, MCP servers, sub-agents, and a 7-level permission model), CC Switch as a universal remote for AI CLIs, MMX-CLI built for agent consumption with clean stdout, semantic exit codes, and async mode, and ConardLi's open-source Garden Skills. A three-level maturity model helps teams progress from personal setup to enterprise deployment. The video production pipeline uses a Vite+React+TypeScript web stage at 1920x1080 instead of generative video models, enabling precise control over typography, timing, and layout. A four-phase workflow with three hard checkpoints ensures quality. The post concludes that the future of agent tooling is designed for digital agents rather than human users.

Key points

What Harness Engineering actually is

A systematic methodology, not a single tool, for making AI agents stable, controllable, and reproducible.
Models already have the capability; the problem is consistency, cross-project reliability, and removing dependence on luck.
A mature Harness consists of six components: context management, tool systems, execution orchestration, state and memory, evaluation and observability, and constraints with recovery.

The core toolchain

Claude Code acts as the flagship carrier, evolved into a multi-agent development environment with Plan Mode (Shift+Tab), deterministic Hooks, Skills via SKILL.md with progressive disclosure (100 tokens scanned, 5K tokens activated), MCP servers for external tools, parallel sub-agents (up to 5), and a 7-level permission model supporting ask, default, and allow modes with deny lists.
CC Switch is a universal remote for AI CLIs (Claude Code, Codex, Gemini CLI) built on React 18 + TypeScript + Vite + TailwindCSS + shadcn/ui + Tauri 2.8 + Rust + SQLite. It offers one-click provider switching, unified MCP management, a prompts and skills marketplace, cross-tool session search, and proxy plus speed-test routing. Removal is non-destructive; data lives in ~/.cc-switch/ with 10 rolling backups.
MMX-CLI is a multimodal command-line tool designed for agents: clean stdout, progress to stderr, --quiet and --output json flags, semantic exit codes for auth, arguments, timeout, and network errors, and --async for parallel work. Capabilities include text chat, image generation, async video generation, 30+ voice TTS, music creation, image understanding, and web search.
Garden Skills (ConardLi) is an open-source Skill collection including web-design-engineer (OKLCH color theory, 520+ advanced patterns), kb-retriever (local knowledge base with PDF and Excel support), gpt-image-2 (18 categories, 80+ structured prompt templates), and web-video-presentation.

Harness maturity model

Level 1 (Personal): 1-2 hours. CLAUDE.md, basic permissions, basic Hooks.
Level 2 (Team): 1-2 days. Shared settings, team agents, MCP, Skills.
Level 3 (Organization): Several weeks. Enterprise MDM, multi-agent orchestration, worktree isolation, entropy management.

How the video production works

Approach: a Vite + React + TypeScript web page at a fixed 16:9 stage (1920x1080 with transform scale), not a generative video model.
A global step cursor driven by (chapter, step) advances via click or keyboard and persists locally.
Each beat owns the full screen; no bullet stacking. A 10-rule design checklist governs typography, color tokens, animation, and reveal pacing.
Four phases with three hard checkpoints: content drafting, user confirmation of script and outline, presentation build, audio synthesis, then recording.
Playback modes: manual (click-to-advance), ?audio=1 (manual advance with auto audio per step), and ?auto=1 (press space once, then auto-advance by audio duration; pairs with OBS for one-take recording).
Audio synthesis: npm run extract-narrations and npm run synthesize-audio call mmx-cli serially with incremental skip and a duration sanity check.
Why not Sora, Runway, Pika, NotebookLM, or Remotion: precise control over fonts, colors, dwell time, and exact numbers; far more stable than model sampling; static-image output from NotebookLM; Remotion constrains the underlying model.

End-to-end collaboration example

web-video-presentation

?auto=1

Key insights

Harness is discipline, not a tool: bound the space, manage state, and place checkpoints.
Web-based video offers roughly 10x more control and stability than sampling a generative video model.
Checkpoints (script, outline, pre-audio) are the single most important guardrail against drift.
Skills package taste, workflow, and quality standards into reusable SKILL.md files.
Agent tooling is shifting from human-facing to agent-facing: clean stdout, semantic exit codes, and async modes are the new baseline.