Module 5.2 · Topic 1

The Text-to-Action Paradigm

Bottom Line: Natural language is executable. By describing outcomes with sufficient clarity—specifying objectives, constraints, quality standards, and success criteria—you enable AI agents to take correct action without…

1.1 From Typing Queries to Describing Outcomes

The difference between a query and an outcome specification determines whether an AI agent executes correctly the first time or returns unusable work. You need to craft instructions that describe desired outcomes with sufficient clarity that an agent executes them correctly without requesting clarification.

Logic behind this approach:

Queries are open-ended: "What is revenue forecasting?" Outcome specifications are closed: "Generate a three-year revenue forecast for a B2B SaaS company with historical Q1–Q3 data, assuming 8% quarterly growth, and present results in a table with monthly breakdowns and confidence intervals." The second instruction tells the agent exactly what success looks like. Precision eliminates back-and-forth.

Sample prompt:

Create a stakeholder communication draft for a proposed operational efficiency project. The project reduces manual approval steps from 7 to 3, affecting 45 employees across three departments. The audience is C-level executives unfamiliar with the specific workflows. Include: (1) a one-paragraph summary of the change, (2) three quantified benefits (time savings, cost, quality improvements), (3) one anticipated concern and mitigation, and (4) a call to action. Use professional but accessible language. Length: 300–400 words.

What to expect in reply:

The agent returns a draft that addresses each component in the exact order specified, with quantified metrics and stakeholder framing. If the generation includes all four elements, specific numbers, and professional tone, the output is usable. If a section is missing or vague, you know the instruction needs revision.

1.2 How Natural Language Becomes Executable Action

You need to predict what actions an agent will take given a description and adjust your instructions if the predicted behavior diverges from your intent.

Translate intent to specification: Start with your goal ("I need a marketing email"). Translate it to an actionable specification that includes role (marketing manager), constraints (mobile-friendly, under 200 words, links to three assets), and quality criteria (conversational tone, no jargon, clear call-to-action). Write the specification so specifically that you can test the agent's compliance.
Submit the specification as an instruction: Provide the full specification to the agent. Include context (target audience is small business owners), constraints (company branding guidelines forbid all-caps headlines), and stopping conditions (do not include pricing; do not solicit email signup).
Review and predict execution: Read the agent's output and trace through: Did it follow the role? Did it respect all constraints? Is the stopping condition honored? If you find a gap (the email is 280 words when you specified 200), note whether the deviation materially affects usability or signals an instruction defect.
Iterate or accept: If the deviation is material, revise the instruction and re-submit. If acceptable, use the output. Pattern your next instruction using this one as a template.

Unlock the Full AI Skill Building Experience

Visit LawQi for access to the full AI Skill Building experience.

Visit LawQi

40% discount using code REASONABLE for personal subscriptions.

Free 48-hour preview access when investigating for teams.