Prompt Kit
Whoever Defines the Work Primitive Wins Prompt Kit
Prompt Kit: Access Is Not Meaning — The Semantic Moat in AI
Every AI product demo this year will look like progress. This kit gives you the frameworks to tell which announcements represent real semantic depth and which are just access wearing a tuxedo. Five prompts, each targeting a distinct decision: evaluating products, auditing your own software, diagnosing agent failures, designing trust architectures, and mapping strategic positioning.
How to use this kit
Each prompt is independent — use whichever one matches your situation. They work best in a thinking-capable model like ChatGPT, Claude, or Gemini, where the AI can reason through layered analysis before producing output. You don't need to fill in any blanks. Every prompt opens by asking you for the context it needs, then runs the analysis. If you're evaluating multiple products, run Prompt 1 once for each and compare. If you're building software, start with Prompt 2 and use the output to inform Prompt 4. If an agent just broke something in production, go straight to Prompt 3.
Prompt 1: The Better Product Test
Job: Evaluate any AI product announcement or demo using the access-vs.-meaning framework from the article — and get a clear verdict on whether it exposes real work primitives or just wraps computer use in spectacle.
When to use: An AI product launches, a vendor sends you a demo, a competitor announces something flashy, or your team is debating whether to adopt a new agentic tool.
What you'll get: A structured evaluation covering what the product actually exposed (new actions, permissions, risk classes, validation paths), where it sits on the access-to-meaning spectrum, what's missing, and a direct recommendation on whether the product is strategically durable or demo-deep.
What the AI will ask you: What product or announcement you want to evaluate, and any supporting material you have (demo link, blog post, documentation, press release, or just your description of what you saw).
Prompt 2: Agent-Readiness Audit
Job: Assess how agent-native your software product actually is — mapping the gap between what humans see in the UI and what an agent can structurally understand — then produce a prioritized roadmap for semantic exposure.
When to use: You're building software and want to know how ready it is for a world where agents (not just humans) interact with it. Or you're a product leader deciding where to invest in agent-facing interfaces versus just adding a chat pane.
What you'll get: An inventory of your product's work primitives, an honest assessment of current semantic exposure, a gap analysis, and a prioritized roadmap for making the product agent-native — not just "AI-enabled."
What the AI will ask you: What your software product does, what domain it serves, what kinds of actions users take, and what interfaces currently exist (APIs, integrations, UI-only workflows).
Prompt 3: Agent Failure Diagnosis
Job: When an agent gets the action right but the decision wrong — or breaks something that looked fine in testing — diagnose whether the root cause was an access problem or a meaning problem, identify which specific type of meaning was missing, and recommend the structural fix.
When to use: An agent misbehaved in production or testing. Examples: a marketing AI went off-brand, a coding agent shipped a confident but wrong fix, a support agent issued a refund it shouldn't have, a scheduling agent broke a politically sensitive meeting, a procurement agent approved a vendor outside policy.
What you'll get: A root-cause analysis that goes beyond "the AI made a mistake" to identify the specific semantic gap — was it missing object awareness, permission context, risk classification, consequence understanding, or validation? Plus a structural recommendation to prevent recurrence.
What the AI will ask you: What the agent did, what it should have done, and as much context as you have about the system setup, permissions, and what information the agent had access to.
Prompt 4: Trust Architecture Designer
Job: Design a scoped authority model for an agent deployment — mapping every action class to its appropriate permission level, review requirement, and escalation path. Turn "trusted write access" from a single switch into a graduated architecture.
When to use: You're deploying an agent into a real workflow and need to decide what it can do autonomously, what needs human approval, and what it shouldn't touch at all. Or you're designing the permission model for an agentic product.
What you'll get: A complete trust architecture: action taxonomy with permission tiers, review requirements for each tier, escalation rules, a rollback plan, and the specific conditions under which autonomy can safely expand over time.
What the AI will ask you: What domain the agent operates in, what actions it needs to perform, who the stakeholders are, what your risk tolerance is, and what approval structures already exist.
Prompt 5: Semantic Moat Analyzer
Job: Evaluate where a company sits on the access-to-meaning spectrum and whether its strategic position is building a durable platform or becoming a feature — using the Salesforce-vs.-SAP and Stripe-vs.-checkout-clickers framing from the article.
When to use: You're making an investment decision, a partnership decision, a build-vs.-buy decision, or a competitive strategy decision about a company in the AI ecosystem. You need to understand whether its position strengthens or erodes as agents get more capable.
What you'll get: A strategic positioning analysis covering what semantic layer the company owns (if any), how defensible that layer is, what happens to its position as agents improve, and where it's vulnerable to being reduced to infrastructure behind someone else's agentic interface.
What the AI will ask you: What company or product you want to analyze, what you already know about its positioning, and what decision you're trying to inform.