Agentic Harness Explainer Script

[TIMING: 3-4 minutes total]
[VISUAL STYLE: Architectural diagrams, system flows]
[TONE: Technical storytelling, accessible but substantive]

Act 1: The Hook

[0:00-0:20]
[DIAGRAM: title-card.svg]

What if building software with AI wasn't about typing prompts... but orchestrating a factory?

[PAUSE 2s]

You've used AI assistants. You've had great sessions. You've also had sessions where everything went sideways. Same tool. Different outcomes. The difference isn't the AI - it's whether you're improvising or producing.

Act 2: The Problem

[0:20-1:00]
[DIAGRAM: problem-flow.svg - Shows scattered prompts, inconsistent outputs]

Here's the pattern most teams are stuck in:

Someone has a task. They open Claude. They type a prompt - maybe a good one, maybe not. They get output. Maybe it's useful. Maybe they spend an hour fixing it. Tomorrow, someone else has a similar task. They start from scratch. No memory of what worked. No accumulated knowledge. No quality control.

[PAUSE 1s]

This is ad-hoc prompting. And it has three problems:

Results are inconsistent. Same request, wildly different quality.
Context doesn't persist. Every session starts at zero.
Quality doesn't compound. Good prompts get lost. Bad patterns repeat.

The ceiling is low because nothing accumulates.

Act 3: The Solution Architecture

[1:00-2:00]
[DIAGRAM: system-architecture.svg - Factory metaphor visualization]

The agentic harness is a different model entirely.

Instead of ad-hoc conversations, you have a factory. The factory has defined inputs - context that persists, skills that encode expertise, memory that carries forward. The factory has assembly lines - workflows that standardize how work moves through stages. The factory has quality control - gates that check work before it proceeds.

[PAUSE 1s]

The key terms:

Agents - specialized configurations optimized for different stages. A planner that researches before committing. An executor that implements without drifting. A verifier that checks work against criteria.

Gates - explicit checkpoints work must pass. Not "looks good" - specific, verifiable conditions.

Phases - structured progression through stages. Research before planning. Planning before execution. Verification before completion.

Memory - context that persists between sessions. What worked. What didn't. Decisions made. Knowledge accumulated.

[PAUSE 2s]

The factory doesn't make AI smarter. It makes AI predictable. Consistent. Improvable.

Act 4: Component Structure

[2:00-2:40]
[DIAGRAM: component-layers.svg - Three stacked layers]

The harness works through three layers:

Constitution - Who Claude is in your context. Your preferences, your standards, your domain knowledge. Loaded automatically. Always present.

Skills - What Claude can do when invoked. Reusable capabilities that encode expertise. Code review. Documentation. Test generation. Each one works the same way every time.

Commands - What Claude does right now. Specific operations that invoke skills with task-specific context.

[PAUSE 1s]

Constitution provides identity. Skills provide capability. Commands provide action. Mixing them creates confusion. Separating them creates clarity.

Act 5: Where to Start

[2:40-3:00]
[DIAGRAM: getting-started.svg - Clear first steps]

If you're evaluating whether this fits your organization:

Ask: Are we getting inconsistent results from AI tools? Is quality unpredictable? Do we restart from scratch each session?

If yes - the harness addresses those problems directly.

[PAUSE 1s]

If you're ready to advocate internally:

The pitch is simple: infrastructure for AI development, not just AI tools. The teams that build infrastructure ship faster, at higher quality, with compounding advantage.

[PAUSE 1s]

If you're ready to start building:

Begin with constitution - a CLAUDE.md file capturing your context. Add one skill - the pattern you repeat most often. Then layer. Each piece makes everything else work better.

The factory isn't built overnight. But every piece you add raises the floor for everything that follows.

[END]

Production Notes

Diagram Requirements:

title-card.svg - Opening visual, "Building with AI" theme
problem-flow.svg - Scattered prompts leading to inconsistent outputs
system-architecture.svg - Factory metaphor: inputs, assembly lines, QC, outputs
component-layers.svg - Three stacked layers (Constitution, Skills, Commands)
getting-started.svg - Clear first steps visualization

Narration Style:

Conversational but authoritative
Technical terms introduced with immediate explanation
Pauses for visual transitions
Builds from problem to solution to action

Target Audience:

Technical leaders evaluating AI tooling
Practitioners curious about structured approaches
Anyone frustrated with ad-hoc AI results