Verify course content

This is an audit, not a sync: it hunts factual drift between the course and the platform, writes a ledger of findings, and changes no lesson. You decide the fixes.

The course lives in this repo (lys-docs): five tracks — customer, agentowner, skillanbieter (persona tracks) and n8n, claude-code (builder tracks) — each a set of de/ and en/ lessons plus MISSION.md and index.html. The source of truth is the platform's internal docs at ~/workspace/linkyourskill.ai/docs. Sibling teaching skill: lys-teach.

Source of truth & precedence

The whole audit turns on this. Read it before assigning any verdict.

• The platform docs are the reference. A DE claim they refute is COURSE-WRONG.
• The docs are not infallible — they lag production too. When a doc contradicts the live-verified ground-truth memory [[lys-platform-ground-truth]], that gap is a SOURCE-CONFLICT: record it and leave the course copy alone. Never rewrite correct course copy to match a stale doc.
• A lesson that teaches around a known production bug is correct pedagogy — the docs' happy-path is the SOURCE-CONFLICT, not the lesson. Known bugs live in [[lys-platform-ground-truth]]; the standing three are literal-only search_skillanbieter, messenger approval delivery broken on prod, and token rotation orphaning old orders. Recall that memory fresh each run — production drifts.
• Count-free copy is correct even when a doc cites a number. The tool-catalog count drifts; an earlier hardcoded "18" already went stale.

The audit

1. Assemble the audit set. Name every in-scope DE file: the lessons under de/lessons/ for all five tracks, each track's de/MISSION.md and de/index.html, and top-level GLOSSARY.md, RESOURCES.md, README.md, index.html. Done when the list is written out and every in-scope DE file appears in it by name.

2. Load the source of truth. Read ~/workspace/linkyourskill.ai/docs/index.md for the doc map, and recall [[lys-platform-ground-truth]] for the known docs↔reality conflicts. Done when you can name which docs cover each track (see routing map) and hold the current known-conflict list.

3. Audit each track against the docs — fan out. Spawn one read-only reader per track (five, in parallel — inline subagents via the Agent tool, not a billed Workflow unless the user asks). Each reader, for its track: extracts every checkable claim from the DE files, cross-references each against the mapped docs, assigns a verdict + severity, and checks whether each EN twin states the same facts as its DE original (drift). It returns a per-track ledger. Done when every checkable claim in every DE file carries a verdict and every EN lesson has a drift result — no file unread, no claim unjudged.

4. Consolidate. Merge the five ledgers into one. A claim that recurs across tracks (the MCP endpoint, search behavior) collapses to a single row listing all its sites. Done when one ledger holds every finding, each traceable to its file.

5. Report. Write ./CONTENT-AUDIT.md: a ledger table — file · course says · docs say · verdict · severity — grouped by severity then track, above a one-line run header (sources, date, file count). Leave every lesson file exactly as found. Then summarize the high-severity COURSE-WRONG and SOURCE-CONFLICT rows inline in chat. Done when CONTENT-AUDIT.md holds every finding, names no secret, and the inline summary lists the high-sev items.

Verdicts

Per claim, one of: OK (correct — omit from the report), COURSE-WRONG (docs refute it, and docs agree with ground-truth), SOURCE-CONFLICT (docs and ground-truth disagree — course untouched), MISLEADING-OMISSION (a taught topic drops a caveat that would trip a learner — e.g. teaching search without its literal-match behavior, or a finale resting on the broken approval delivery), UNVERIFIABLE (no doc speaks to it). Severity: high / med / low.

What is a checkable claim

Anything a doc can confirm or refute: the MCP endpoint and tool names/args, required order fields, the order state machine and events, search behavior, prices, auth/token wire shapes, the messenger/approval flow, glossary definitions. Narrative, motivation, and tone are out of scope — pass over them.

Doc routing map

Course area	Primary docs
n8n, claude-code (builder)	mcp-guide.md, api-reference.md, api-parity-matrix.md, architecture.md, internal-api-contracts.md, messenger-adapter.md
customer	journeys/customer.md, architecture.md, api-reference.md
agentowner	journeys/agentowner.md, messenger-adapter.md, architecture.md, security.md
skillanbieter	journeys/skillanbieter.md, ai-domain-skill-standards.md, database.md
GLOSSARY.md	architecture.md, database.md, journeys/*
RESOURCES.md	mcp-guide.md, deployment.md, development-setup.md

Guardrails

• Report-only. The deliverable is the ledger; every lesson file stays exactly as found.
• Leak-clean. The report names no token, workspace/connector ID, account email, or staging host — the same leak rule that binds the course. See [[credential-handling-wrap-proof]].