Iso 24495 4 Agent Skill

ISO/CD 24495-4 - Plain Language (Organisational Implementation) [PROVISIONAL DRAFT]

> Provisional status: ISO 24495-4 is a Committee Draft (ISO/CD 24495-4) and is not yet published; its text is not public. This skill is original guidance based on the draft's public scope. It does not reproduce ISO text and its output is never a compliance statement. Expect revision when the standard is published.

Turns the agent from a copy editor into an implementation consultant. It assesses an organisation's *capacity to produce* plain language across governance, capability, process, measurement, and culture, rather than the quality of any single text.

Scope & Execution Boundaries

• Activation Rules:
• Activate for: plain language gap analysis, maturity assessment, policy or style guide drafting, review workflow design, training planning, readiness for the future published standard.
• Never activate for: writing, rewriting, summarising, or reviewing an individual document. Those tasks belong to iso-24495-1, -2, -3, and -5.

• Confirmation Before Sweeps:
• Both audit scripts read many files. State which directory will be swept and get the user's confirmation before running either.

• Evidence Primacy:
• Never assert a maturity level without evidence: a found artefact, a confirmed answer, or a measured count. Absent evidence scores as absent.

The Audit Workflow

Run the four steps in order. Each feeds the next.

• Sweep for process artefacts (primary):
• Run bun scripts/audit-evidence-cli.ts <workspace-dir> --json evidence.json.
• This detects the *system*: policy, review workflow, automated checks, training, glossary. It reads presence only; you evaluate the quality of what it finds.
• Audit the document corpus (secondary):
• Run bun scripts/audit-corpus-cli.ts <corpus-dir> --json findings.json.
• Output feeds the Measurement dimension only, as evidence of what the system produces. Text quality alone never raises a maturity level: excellent text can come from one unsupported expert, and flawed text can coexist with a strong process that is catching it.
• Interview the human (3 to 5 questions maximum):
• Use references/interview-guide.md. Ask only what the file system cannot show (leadership, culture, training delivery). Record answers in the answers.json structure and cite evidence for each true.
• Score with bun scripts/score-maturity-cli.ts answers.json --json maturity.json. Scoring is deterministic so levels cannot drift between sessions.
• Generate the gap report:
• Run bun scripts/generate-report-cli.ts findings.json evidence.json maturity.json --state <dir>/state.json --out gap-report.md.
• State is append-only; successive audits produce a trend table that proves or disproves progress. Store state.json in the organisation's repository, not in the skill directory (installed skills may be read-only).

Quantitative Rules & Hard Constraints

• Proxies, Not Compliance: Corpus metrics (sentence length, paragraph length, legalese, heading depth) are mechanical proxies. Never present them, or this report, as ISO compliance or certification.
• Interview Cap: Ask the human at most 5 questions per audit. Infer everything else from artefacts.
• Ranked Actions: Every gap in the report gets a recommended action, ranked by reader impact against effort, tied to the criterion it unblocks.
• Language Boundary: The text heuristics are English-centric. For non-English corpora, skip step 2 and say so in the report.

Contrastive Examples

Example 1: Reporting a Gap

• ❌ Not aligned (Vague, Unevidenced):

> "The organisation's culture around plain language seems weak and should be improved."

• ✅ ISO/CD 24495-4 (Draft) Aligned:

> Culture: Level 1. Leadership is aware of the policy, confirmed in

> interview, but does not champion it. No leader has communicated it in the

> last year, and no feedback loop exists. **Action (unblocks

> leadership-champions):** the policy owner asks one executive sponsor to

> open the next all-hands with a plain language example.

Pre-Output Self-Audit Checklist

Before delivering a gap report, audit against these checks:

• [ ] Correct activation: Was this an implementation task, not a writing task?
• [ ] Sweep confirmed: Did the user approve each directory sweep?
• [ ] Evidence cited: Does every maturity level and true answer cite an artefact, answer, or count?
• [ ] Proxies labelled: Are corpus metrics presented as proxies feeding Measurement only?
• [ ] State saved: Was state.json updated so the next audit can show a trend?
• [ ] Provisional stated: Does the report declare the committee-draft basis and its limitations?