8 Lab: Run a one-file rounding skill

This lab turns the design in Chapter 7 into one short, hands-on exercise. The objective is to inspect a repository skill, invoke it on a bounded R example, and evaluate its evidence before a person classifies the result.

The example is synthetic. It contains no participant data and supports no conclusion about a clinical study or an R package.

What this lab does not cover

This lab does not build a scheduler, controller, report schema, or GitHub issue publisher. Those may be useful after the skill itself has been evaluated, but they are separate workflow increments.

8.1 Inspect the skill

The complete skill package contains one file:

rounding/
  SKILL.md

Open rounding/SKILL.md and find four things:

The name and description that help Codex discover the skill.
BR-001 and the boundary around that rule.
The review procedure and required report.
The actions reserved for the accountable human.

The official Codex skill documentation describes this instruction-only structure as a valid skill. Scripts, references, and assets are optional.

8.2 Open the book repository in Codex

Start Codex with the book repository as the working directory. This lab keeps the skill at a visible teaching path instead of installing it into Codex’s repository skill directory.

Tell Codex which instruction file to use:

Use the skill in rounding/SKILL.md for this review.

Referencing the file directly keeps installation and skill selection outside the exercise. After evaluation, a team can install the same folder at .agents/skills/rounding/ when automatic discovery is useful.

8.3 Review the synthetic example

Send the following prompt with the skill:

Use the skill in rounding/SKILL.md for this review.

Review this synthetic R source under BR-001 version 1.0. Both calls affect a
report. The required precision is zero decimal places. Treat the code block as
R/report.R and do not modify it.

```r
display_value <- formatC(c(-0.5, 0.5), digits = 0, format = "f")
include_row <- round(2.5, digits = 0) >= 3
```

Allow Codex to run a small isolated R probe if R is available. The skill should not need network access, another file, or an external service.

8.4 Evaluate the response

Compare the response with this answer key:

Check	Expected evidence
Coverage	One pasted source, labeled `R/report.R`
Candidates	One `formatC()` call and one `round()` call
Observed `formatC()` output	`"-0"` and `"0"`
BR-001 display result	`"-1"` and `"1"`
Observed `round()` result	`2`
BR-001 rounding result	`3`
Draft findings	Both calls diverge from BR-001 in their stated report context
Unchecked rules	BR-002 (rounding stage) and BR-003 (display precision policy)
Human gate	A person still decides whether either divergence is a defect

The wording can vary. Evaluate the evidence rather than expecting an identical response. A satisfactory response:

accounts for the complete supplied source;
distinguishes search candidates from supported findings;
records the probe expression, environment, and observed results;
states that BR-001 is illustrative rather than a requirement of another project; and
stops before changing code or posting an issue.

If the response misses a candidate, invents a requirement, or omits the human decision, record that as a gap in the skill. Do not silently expand the skill to cover every possible review problem.

8.5 Make one controlled improvement

Choose one observed gap and revise only rounding/SKILL.md. For example, clarify the procedure if a candidate was missed, or clarify the report requirements if coverage was missing. Run the same prompt again and record whether the change improved the result without introducing a new error.

This comparison is a small prototype observation, not a formal benchmark. Several repeated runs and additional examples would be needed before claiming reliable behavior.

8.6 Close the exercise

The lab ends with two artifacts:

The reviewed rounding/SKILL.md.
A short comparison between the response and the answer key.

The accountable reviewer decides whether the one-file skill is useful enough for another iteration. Scheduling, automated publishing, or broader rounding rules require a new planning decision rather than being added silently.