QC skills guide • airsetup

airsetup provides five skills for instructing an AI agent to perform quality control (QC) at different stages of an AI-assisted R workflow. These skills review analysis requests, clinical-trial documents, analysis plans, and R outputs. They produce structured evidence that helps the user decide whether enough information is available to proceed.

QC does not delegate decisions to the AI. It makes uncertainties, evidence, assumption risks, and required actions visible so that the user or responsible reviewer can make an informed decision.

Japanese version

Choose the skill that matches the current stage

Select the skill closest to the current stage of work.

Current situation	Skill	Primary question
An analysis request or data information is available	CONTEXT QC	Is there enough information to start the analysis?
A general analysis plan has been drafted	PLAN QC	Can R code be written from this plan?
A clinical-trial SAP has been drafted	SAP QC	Are the statistical specifications and supporting evidence sufficient?
Information must be organized across clinical-trial documents	M11 SEMANTIC QC	What can be extracted with document-supported evidence?
Tables, figures, or model outputs are available	RESULT QC	Can these results be safely reported and interpreted?

When uncertain, start with CONTEXT QC. If several ordinary skills appear applicable, begin with the earliest unresolved workflow stage.

Context
    ↓
Analysis plan or SAP
    ↓
R coding and analysis execution
    ↓
Results

Even when results already exist, unresolved meanings of the analysis population or variables may require returning to context review rather than relying on RESULT QC alone.

Add skills to a project

airskill() adds Markdown skill files for an AI agent to a project. Creating the files does not itself start QC. After creation, identify the skill to use in the request to the AI agent.

Add all skills:

airskill("project folder")

Add selected skills:

airskill(
  path = "project folder",
  skills = c("context", "plan", "result")
)

Add the clinical-trial skills:

airskill(
  path = "project folder",
  skills = c("sap", "m11_semantic")
)

Skill files are created in ai_project/agent_control/.

Value in `skills`	File created
`context`	`QC_SKILL_CONTEXT.md`
`plan`	`QC_SKILL_PLAN.md`
`sap`	`QC_SKILL_SAP.md`
`m11_semantic`	`QC_SKILL_M11SEMANTIC.md`
`result`	`QC_SKILL_RESULT.md`

AGENT_CONTROL_INDEX.md, created in the same folder, summarizes the purpose and selection rules for each skill.

PLAN QC versus SAP QC

PLAN QC and SAP QC are alternatives selected according to the document under review; they are not normally consecutive mandatory steps.

PLAN QC applies to general analysis plans and coding specifications, including descriptive analyses, cross-tabulations, epidemiologic or registry studies, and exploratory analyses.

SAP QC applies to statistical analysis plans for clinical trials. In addition to implementation readiness, it reviews consistency with the protocol, estimands, analysis sets, interim analyses, multiplicity, missing data, sensitivity analyses, safety analyses, version control, and revision history.

When a clinical-trial SAP is available, select SAP QC rather than PLAN QC in most cases.

When to use M11 SEMANTIC QC

M11 SEMANTIC QC is not mandatory for every clinical trial. It is useful when:

information is distributed across the protocol, SAP, CRF, and data definitions;
estimands or intercurrent events must be organized;
endpoints must be mapped to source variables;
cross-document inconsistencies must be identified;
trial semantics must be organized before SAP development; or
ICH M11 is being used as a reference framework for semantic organization.

For a simple dataset, variable list, or R coding request, CONTEXT QC is usually sufficient. M11 SEMANTIC QC does not certify ICH M11 compliance; it uses M11 concepts as a reference framework for organizing clinical-trial semantics.

Standard workflows

General analysis

Analysis request and data information
    ↓
CONTEXT QC
    ↓
General analysis plan
    ↓
PLAN QC
    ↓
R coding and analysis execution
    ↓
RESULT QC
    ↓
Reporting and interpretation

Clinical trial

Protocol, CRF, data definitions, and related materials
    ↓
M11 SEMANTIC QC, when needed
    ↓
SAP
    ↓
SAP QC
    ↓
R coding and analysis execution
    ↓
RESULT QC
    ↓
Reporting and statistical review

Principles shared by all QC skills

Do not infer missing study-specific information

If information cannot be verified from the documents or data, record it as Cannot assess, Unknown, or an unresolved item rather than filling it with a plausible value.

Separate documented facts from AI proposals

Distinguish document-supported facts, cross-document inconsistencies, and AI proposals. An AI proposal is not an approved study specification.

Provide evidence

Where possible, cite filenames, document sections, tables, variable names, code, or logs that support each judgment.

Do not reduce QC to a score

Record the issue, impact, evidence, and recommended action. A single critical issue may block the next step even when most other items appear complete.

Interpreting QC results

Status	Meaning
`OK`	No problem was identified within the provided materials
`Needs clarification`	Confirmation by the user or responsible reviewer is required
`Problem`	Correction or another action is required
`Cannot assess`	Required materials are unavailable
`Not applicable`	The item does not apply to the current task

Severity	Meaning
`Critical`	Do not proceed until the issue is resolved
`Major`	Resolve before the result or implementation is used
`Minor`	Correction is desirable but normally does not change the main conclusion
`Note`	Limitation, caution, or reference information

Status and severity are separate. For example, an analysis-population definition may be Cannot assess and still constitute a blocking issue.

Requesting QC from an AI agent

Name the skill

Use QC_SKILL_CONTEXT.md to perform CONTEXT QC
for the current analysis request.

Do not infer missing information, and explicitly identify
items that cannot be assessed.

Ask the AI agent to select a skill

Review the available materials and the current workflow stage,
then select and apply the most appropriate airsetup QC skill.

State which skill you selected and why before starting the review.

Ordinarily, do not run several skills simultaneously. Start with the earliest unresolved stage.

After receiving a QC report

Use the report as a working document rather than as a certificate of approval.

Review Critical and Major issues.
Identify Cannot assess items that affect the next step.
Answer clarification questions and decision items.
Revise the documents, data definitions, plan, code, or output.
If needed, conduct another review using the next sequential report number.
Carry unresolved items into QC_STATUS.md.

Do not normally overwrite earlier reports; retain them so the review history remains traceable.

Detailed references

The scope of each QC skill depends on the supplied materials. These skills do not automatically guarantee the accuracy of source data, code, statistical methods, or scientific conclusions. Important decisions require review by the appropriate investigator, statistician, data specialist, or other responsible person.