Trace Human Intent to Machine Evidence with human-intent-audit

Where this fits

You ask an agent to "make the dashboard load faster," and it reports a 200ms improvement in server response time. But your actual goal was perceived performance on slow mobile networks—something the agent never measured. This skill, human-intent-audit, plugs that gap by forcing agents to map every natural-language requirement to concrete observables, evidence, and acceptance criteria that a human owner can review.

Why agents benefit

• Distinguishes machine-green from human-pass: The skill enforces that a passing test or metric can't be claimed as success unless it has a proven semantic link to the user's final outcome.
• Auto-generates evidence chains: It walks the agent through building a trace from intent → machine predicates → observables → evidence, so no step is skipped.
• Prevents metric invention: Agents stop guessing thresholds like "fast" or "safe"—they either derive them from context or ask the user for a genuine product choice.
• Produces owner-readable acceptance: The final audit output is structured for humans, not just machine logs, making it easy for stakeholders to verify the work.
• Reduces rework and false confidence: By catching semantic gaps early, the skill avoids the classic failure of shipping something that passes tests but fails the user's actual need.

Practical scenarios

Verifying a UI change matches "smooth scrolling"

An agent modifies scroll behavior and reports a frame rate increase. With human-intent-audit, it must define what "smooth" means to the user (e.g., no jank on 60Hz displays), measure actual frame timings, and show a side-by-side comparison that a product manager can approve.

Auditing a security fix for "safe"

You ask an agent to patch a vulnerability and "make it safe." The agent can't just run a scanner and say PASS. The skill forces it to define the threat model, enumerate attack vectors, and provide evidence that each one is mitigated—plus a clear statement of residual risk.

Checking data migration for "all records preserved"

The agent migrates a database and reports zero errors. But "all" might mean every row, every relationship, or every field—even nulls. human-intent-audit makes the agent clarify the scope, implement row-by-row verification, and produce a reconciliation report with counts and exceptions.

Add it to your agent workflow

Install the skill via npx:

npx skills add ConnorRX56/presentation-delivery-skills --skill human-intent-audit

Then invoke it in any audit or completion gate. For example, after an agent finishes a task, prompt:

Run human-intent-audit against the original request and my acceptance criteria.
Produce a traceability report with evidence for each requirement.

The agent will generate a structured audit, flag any unproven links, and ask you for missing definitions instead of guessing.

Compare before adopting

Before committing, evaluate the skill against your workflow: check if the quality signals (e.g., rule enforcement, evidence format) match your team's review process. Look at maintenance freshness—is the repo active? Compare with alternatives like generic checklists or custom prompt templates. Consider whether the skill's strictness fits your tolerance for agent autonomy or if you need a lighter-weight approach for quick tasks.

Why it is worth tracking

With 10 stars and a focused rule set, human-intent-audit is early but addresses a critical failure mode in agent development: the gap between what users ask and what machines verify. The community momentum is small but relevant—security and compliance teams will find this valuable. Track it now to see if the author iterates on real-world feedback; when your team starts doing serious QA or parity checks, this skill could become a standard gate in your pipeline.