Reference: every artifact this course draws on

Quant and Veridian tag every material claim [Established], [Directional], or [Speculative] because they synthesize outside sources of uneven reliability — a peer-reviewed paper next to a practitioner heuristic next to a marketing page. This course doesn't need that system for most of what it says, because most of what it says is a direct quote from a file already sitting in this repository, attributed to that file by name, checkable by opening it. The one category that doesn't work that way is named at the end of this page rather than folded in silently.

#Skills — knowledge/skills/

• research-index — the entry point for this repo's whole research library; linked into every project on the machine rather than symlinked per-project, because it's meant to be found before a session starts researching something already answered.
• learning-science — the evidence behind how a lesson, quiz, or scheduler should actually be built.
• course-builder / course-max — the compact and exhaustive pair covered in the Skills module: eighteen rules and a process, versus the full reusable-prompt library and rubric underneath it.
• course-creator — the skill lost on 2026-08-20 for living outside version control, and recovered only by accident; the cautionary example the Skills module is built around.
• agent-orchestration — the source for nearly everything in the Workflows, Subagents, and Honest Limits modules: 23 real workflow runs, what worked, what broke, and the same-model-reviewer-bias finding.
• micro-polish — the systematic pass for the class of bug that survives a feature technically working.
• meta-improvement-loop — the skill that audits whether a review process itself keeps missing the same class of finding, rather than auditing content directly.
• competitive-content-audit and soul-extractor — present in the plugin; not directly quoted in this course.

#Workflow scripts — knowledge/workflows/scripts/, indexed by knowledge/workflows/README.md

The README is the single best entry point in this whole reference list — its own framing states the reason: "where are all the agents and sub-agents and orchestrators we made... an agent is a prompt, and an orchestrator is a script that decides which prompts run, in what order, and what happens to their output." Anyone about to write a first Workflow script should read it before writing a line of code.

• meridian-payments-security-audit.js — the two-skeptic, keep-if-one-survives adversarial audit on checkout and access-control code, cited in the Workflows module for its deliberately lopsided vote rule.
• meridian-review-council.js — the one-verifier, re-derive-the-answer-yourself pattern used on a course's own answer keys; the 2026-08-28 council whose findings are traced through wec13's seed-version history in the Honest Limits module.
• meridian-porting-fixes-and-content-quality.js, port-official-sat-questions.js, meridian-retry-porting-and-audits.js — the three-attempt chain behind the output-token-ceiling failure and its actual fix (write incrementally; don't paste generated content back into the final response).
• verify-production-deploy.js — the re-run-the-checks-yourself variant of adversarial verify, checking a fact about the world rather than re-judging another agent's reasoning.
• llm-council-protocol-review.js — the FINDING_SCHEMA example showing agent()'s structured-output mechanism in the Subagents module.
• meridian-curriculum-build.js, meridian-mega-pass.js, meridian-fork-reconciliation.js — referenced in the README's own table; not quoted directly in this course, worth reading for the parallel-fan-out and read-only-reconciliation shapes they use.

#Articles

• "Twenty-Three Agents and What Broke" — the retrospective this whole course traces back to, and the direct source for the split-research-from-implementation rule, the adversarial-verify rule, and the output-ceiling failure's actual diagnosis.
• "5,486 Videos, and the Bug That Lied" — the source for the Honest Limits module's silent-failure example: a caption fetcher reporting "no captions" for videos that were actually blocked, and the general lesson that a failure mapping onto an innocent-looking existing state won't be caught by testing that things run.

#Key files

• content/skills.ts — this site's own skill registry and the source of the "trigger condition, not a summary" convention quoted throughout the Skills module.
• content/courses.ts — specifically the wec13 entry's seedVersion comment history, read directly for the Honest Limits module's central evidence.
• scripts/build-quant.mjs — the barebones markdown-to-nodes build script this course's own scripts/build-building-with-claude.mjs copies structurally, module for module.
• content/quant/ — the sibling course built the same way this one is: real source documents compiled straight to lesson nodes, no typed lesson/quiz blocks.

#This course itself

content/building-with-claude/ and scripts/build-building-with-claude.mjs are tracked in the same repository they describe, registered in content/courses.ts the same way every other course on this site is — for the exact reason the Skills module's course-creator incident gives: a file that isn't in version control isn't safe, no matter how good its content is.

#What this course doesn't cover

This isn't a general manual for Claude Code, the Claude Agent SDK, or the Model Context Protocol — it's scoped to what has actually been run against this codebase and left a real artifact behind. Skipped entirely: MCP tool authoring, the deeper mechanics of context-window management, and any Claude product surface this repository's own workflows haven't exercised. Where a claim in this course describes how Skills, subagents, or Workflows behave day to day rather than quoting a specific file that states it in one place — the mechanics of description-as-trigger, or of an agent() call's isolation — that claim is drawn from using the tools rather than from one citable line, and is named here as the one category in this course that isn't pinned to a file you can open and check directly.