Sources and research notes

Theo Browne

The lesson titles, dates, and concept lists come from the unchanged supplied course outline. The five teaching texts, Patch Desk example, prompts, exercises, expected results, and answers are independently written. They are not MEGA lesson bodies and are not attributed to Theo. Research was performed on 2026-10-19 ET. No claim of video playback is made.

Instructor material actually read

• What even is an agent harness?: the web reader returned no usable body, but yt-dlp retrieved publicly available English automatic captions. Read the opening and caption windows around 3:30–6:30, 8:20–10:20, and 18:20–20:00. The short summaries in lessons 1 and 4 use the distinction between generation, tool execution, observations, and the next model input. No numerical model-performance claim from the video is reused. Automatic captions can misrecognize product names and words; no direct quote is used.
• Delete your CLAUDE.md (and your AGENT.md too): web reader exposed metadata only; yt-dlp retrieved English automatic captions. Read the opening and windows around 4:30–6:30, 17:30–19:30, and 28:20–30:20. Lessons 2, 3, and 5 use brief paraphrases of the context-file audit, targeted guidance, and codebase/check feedback. The video's discussion of a study was not treated as independent verification of that study and no performance figure was imported.
• Theo's site and blog index: read for first-party discovery. The index is not evidence for agent mechanics.
• T3 Code repository: README read. Supports the limited claim that the application is a control surface for installed agent providers. This is team-maintained product documentation, not Theo's personal statement about every lesson topic.
• T3 Code permission modes: read the mode table, approval behavior, and provider differences. Lesson 4 uses only the existence and purpose of these controls, with a reminder to inspect live settings.
• T3 Code source control: read change-review and source-control workflow material. Lesson 3 refers to inspecting edits, not to a guarantee of publication or correctness.

Supplementary primary technical sources

• Hugging Face Transformer mechanics: read causal prediction, training/fine-tuning, architecture, and attention sections. Supports lesson 1's compact mechanism explanation. The invented Patch Desk reasoning is original.
• Hugging Face generation reference: read configuration/control descriptions. Supports only sampling and output-length distinctions, not a universal recommended temperature or a truth guarantee.
• Hugging Face bias and limitations: read to support the short limitations paragraph in lesson 1.
• Anthropic context engineering: read context composition, runtime retrieval, and note-taking sections. Briefly supports lessons 2 and 3. The packet, file layout, and maintenance policy are original applications.
• Agent Skills specification: read directory layout, metadata, body, and optional resources. Supports the narrow procedure-packaging descriptions in lessons 2 and 3.
• Anthropic tool design: read tool descriptions and response design. Supports the short discussion in lesson 2. No vendor SDK is reproduced.
• Anthropic Building effective agents: read workflow/agent distinction and complexity tradeoffs. Supports lesson 5, not a universal rule against autonomous systems.
• Git worktree reference: read description of multiple working trees. Supports lesson 4's file-isolation explanation. No claim that it isolates external resources is made.
• OWASP prompt-injection prevention: read indirect injection and layered execution controls. Brief support in lessons 2 and 5. The example does not promise a complete injection defense.
• W3C input validation: read accessible error/recovery guidance. Supports one short user-experience paragraph in lesson 5.
• MCP tools specification: read schemas/results material. Supports the interface/result distinction in lesson 5, not an implementation of MCP.

Unavailable material and boundaries

• Supplied video Jf54k7tFeEc: web reader returned a cache miss. Caption retrieval required a bot-confirmation sign-in, so it was not used. No cookies, paid access, or private account access was used to bypass this limit.
• Supplied X posts remain in the original research file. They were not independently fetched for these lessons and are not silently treated as newly verified source evidence.
• TubeReads and other third-party video recaps were not used for lesson claims. Public automatic captions, where available, were read directly instead.
• Fenced task packets and pseudocode are illustrative teaching artifacts. They are not claims of executing the examples or tested integrations.

Lesson mapping

• Lesson 1: video tool-call mechanics plus first-party model documentation; original command-correction example.
• Lesson 2: video context guidance plus context, skills, tools, and injection references; original evidence packet and delegation contract.
• Lesson 3: video context-file audit plus retrieval/skills/source-control references; original authority and document-maintenance design.
• Lesson 4: harness video and first-party control-surface, permissions, and Git references; original failure diagnosis.
• Lesson 5: video feedback guidance and primary workflow, validation, security, and tool-interface references; original state machine and acceptance exercise.

Angie Jones

Research performed during the 2026-10-19 task. Five lessons were independently written for the exact titles and topic lists in data/course.json. The original input data was not edited. These are original educational syntheses, not MEGA's paid lesson bodies, transcripts, instructor-authored companions, or claims about unpublished teaching. Cedar Cycle, its policies, cases, task packets, tools, code, exercises, and expected outcomes are fictional teaching material created for this project.

Method and attribution boundaries

The assigned background author read the research skill and used web browsing to open actual primary pages before citing them. Instructor-authored pages were preferred, supplemented by official protocol/format documentation and first-party engineering articles. No secondary recap was used as evidence. No verbatim quotation from public prose is included in the lesson bodies. Short attributed source descriptions are kept narrow; most prose develops original examples and engineering judgments rather than restating a source's structure. All pseudocode is labeled illustrative and API-neutral; no live API behavior or measured results are claimed.

The two Goose articles on observability and server safety are team publications, not Angie-authored articles. The lesson explicitly attributes the observability article to the Goose team and names its author. Finding a page through Angie's author index does not establish that she wrote every item in the surrounding site navigation.

Sources used in lessons

Agentic AI and the MCP Ecosystem — Angie Jones

URL: https://angiejones.tech/agentic-ai-and-the-mcp-ecosystem/

Read evidence: the opened page contained the agent introduction and the tool-calling example, including model-selected reads, edits, test execution, and returned results (page lines 34–108).

Supports: the basic model/tool/application observation cycle in lesson 1. Historical model names, ecosystem counts, broad integration claims, and the source's code example were not reproduced.

Access: article body available. Lesson mapping: content/angie/1.md.

Agent Memory — Angie Jones

URL: https://angiejones.tech/agent-memory/

Read evidence: the article body supplied named memory categories and discussed maintenance, retrieval, correction, and tenant scoping (lines 49–168). Its Oracle-specific section was visible but is not used as an implementation dependency.

Supports: the brief memory-category attribution in lesson 1. The lesson's records, freshness rules, and support examples are independently designed.

Access: full article body available. Lesson mapping: content/angie/1.md.

3 Prompts to Test for Agent Readiness — Angie Jones

URL: https://angiejones.tech/3-prompts-to-test-for-agent-readiness/

Read evidence: tool calling, follow-up file modification, and reading by contextual reference appear in the opened body (lines 34–92).

Supports: a small multi-turn capability probe in lesson 2. The lesson adds a domain-specific evaluation proposal and deliberately does not treat a three-prompt pass as a production guarantee.

Access: article body available. Lesson mapping: content/angie/2.md.

Building Effective Agents — Anthropic engineering

URL: https://www.anthropic.com/engineering/building-effective-agents

Read evidence: the opened page discusses when agents add value, simple starting points, and framework abstractions (lines 28–49).

Supports: the narrowly attributed simplicity/framework tradeoff in lesson 2. Cedar Cycle's framework acceptance checks and state machine are original design proposals.

Access: article body available. Lesson mapping: content/angie/2.md.

AGENTS.md — project documentation

URL: https://agents.md/

Read evidence: setup/style example and rationale for agent-focused guidance separate from README content (lines 9–45).

Supports: the role of a predictable project instruction file in lesson 3. The Cedar Cycle AGENTS.md excerpt is original, not copied from the site.

Access: documentation body available. Lesson mapping: content/angie/3.md.

Architecture overview — Model Context Protocol

URL: https://modelcontextprotocol.io/docs/learn/architecture

Resolved version during research: https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture

Read evidence: scope, host/client/server roles, local and remote implementations, layers, discovery, and primitives were visible (lines 48–134 and 168–220).

Supports: the short architecture account in lesson 3. Wire details are deliberately omitted because installed clients may implement different versions. The source's current metadata/discovery flow is not presented as a universal historical handshake.

Access: official documentation body available; URL is version-sensitive. Lesson mapping: content/angie/3.md.

Agent Skills Overview — Agent Skills project

URL: https://agentskills.io/home

The initial requested URL, https://agentskills.io/what-are-skills, redirected to this page.

Read evidence: folder structure, SKILL.md metadata/instructions, optional resources, and discovery/activation/execution loading stages (lines 25–55).

Supports: format and progressive disclosure in lesson 3. The billing skill, trigger wording, validator, and file examples are original.

Access: official documentation body available. Lesson mapping: content/angie/3.md.

Did Skills Kill MCP? — Angie Jones, Goose blog

URL: https://goose-docs.ai/blog/2025/12/22/agent-skills-vs-mcp/

Read evidence: title, author byline, and the distinction between workflow organization and executable integrations were visible (lines 118–142 and following article body).

Supports: the complementary roles of skills and integrations in lesson 3. The lesson avoids adopting the article's broad suggestion that most real-world agent capabilities necessarily involve MCP; direct tools and other integration approaches also exist.

Access: article body available. Lesson mapping: content/angie/3.md.

3 Design Principles for Creating Agent Skills — Angie Jones

URL: https://angiejones.tech/3-design-principles-for-creating-agent-skills/

Read evidence: article body and dated byline available; sections distinguish consistent scripted decisions, model judgment, and explicit constraints (lines 65–118).

Supports: a brief design attribution in lesson 3. The original Repo Readiness scoring example, exact rules, and conversation arc are not recreated. The Cedar Cycle validator example is original.

Access: article body available. Lesson mapping: content/angie/3.md.

Agents, Subagents, and Multi Agents: What They Are and When to Use Them — Angie Jones, Goose blog

URL: https://goose-docs.ai/blog/2025/08/14/agent-coordination-patterns/

Read evidence: author byline and topology distinctions appear in lines 119–179. The source's dark-mode example was read but not reused.

Supports: ownership distinction among a single actor, parent-led subagents, and independent collaborating agents in lesson 4.

Access: article body available. Lesson mapping: content/angie/4.md.

Agent Discovery — A2A Protocol

URL: https://a2a-protocol.org/latest/topics/agent-discovery/

Read evidence: Agent Card purpose, capability and interaction information, discovery strategies, and security considerations (lines 103–180).

Supports: the narrow Agent Card/discovery explanation in lesson 4. The fictional warranty description is explicitly not a complete valid Agent Card. No registry API is invented.

Access: official documentation body available; latest-version URL may evolve. Lesson mapping: content/angie/4.md.

Life of a Task — A2A Protocol

URL: https://a2a-protocol.org/latest/topics/life-of-a-task/

Read evidence: immediate messages, stateful tasks, progress/input needs, and terminal/interrupted states (lines 98–112 and subsequent sections).

Supports: distinguishing acknowledgment, ongoing work, and completion in lesson 4. Cedar Cycle's timeout and late-result rules are proposed orchestration policy, not claimed protocol guarantees.

Access: official documentation body available. Lesson mapping: content/angie/4.md.

Test Pyramid for AI Agents — Angie Jones

URL: https://angiejones.tech/test-pyramid-for-ai-agents/

Read evidence: deterministic foundations, recorded/replayed interactions, repeated performance measurements, rubric judgment, and testing tradeoffs (lines 53–130).

Supports: the short separation-of-test-layers attribution in lesson 5. The four Cedar Cycle cases and graders are new. The source's scripts, model judging schedule, and project-specific CI policy are not copied or generalized as requirements.

Access: article body available. Lesson mapping: content/angie/5.md.

Demystifying Evals for AI Agents — Anthropic engineering

URL: https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents

Read evidence: code-based, model-based, and human graders and their tradeoffs (lines 53–120).

Supports: a brief grader-selection account in lesson 5. Cedar Cycle's dimensions and acceptance criteria are independently designed, with no claimed benchmark results.

Access: article body available. Lesson mapping: content/angie/5.md.

How goose Catches AI Errors with Langfuse — Tania Chakraborty, Goose team

URL: https://goose-docs.ai/blog/2025/03/18/goose-langfuse/

Read evidence: author byline and trace/timeline/structured-data descriptions were read (lines 118–155).

Supports: the short trace-based observability attribution in lesson 5. This is a primary product-team article by Tania Chakraborty, not an Angie Jones article. No claims are made from the linked livestream itself.

Access: article body available. Lesson mapping: content/angie/5.md.

Original seed access limitations

• https://block.github.io/goose/blog/authors/angie/ returned 404 through the browsing tool. The available project-hosted replacement, https://goose-docs.ai/blog/authors/angie/, exposed the author index and links used above.
• https://www.youtube.com/watch?v=NtAyOP-zlvo exposed the title “Workshop: Block: Vibe Coding with Goose Building Apps with AI Agents and MCP by Block” plus a minimal YouTube footer. No transcript, substantive description, or video contents were available. No teaching claims were derived from the video.
• https://x.com/techgirl1908/status/2108989477544722640 returned an internal fetch error.
• https://x.com/techgirl1908/status/2106045938821112161 returned an internal fetch error.
• https://x.com/techgirl1908/status/2104973526826725703 returned an internal fetch error.
• https://x.com/techgirl1908/status/2102791576460365972 returned an internal fetch error.
• Those unavailable seed posts remain unchanged in the original input data; none is used to substantiate a lesson claim.

Other opened research, not relied on in lesson bodies

• https://angiejones.tech/category/ai/ — discovery index with links to the first-person posts above; not used as a substitute for reading articles.
• https://angiejones.tech/ — first-person site/homepage; no current-role claim is needed in the lessons.
• https://angiejones.tech/wait-are-we-just-handing-over-system-access-to-the-ai-agents/ — fetch error; not used.
• https://goose-docs.ai/blog/2026/05/05/goose-as-conductor/ — readable project-team orchestration article; not needed for the final source set.
• https://goose-docs.ai/blog/2025/03/26/mcp-security/ — readable Goose team server-safety article by Ebony Louis; not represented as Angie-authored and not cited in final lessons.
• https://github.com/aaif-goose/goose/blob/main/documentation/docs/guides/context-engineering/subagents.mdx — official project source surfaced during discovery; implementation-specific defaults were deliberately not adopted in API-neutral teaching.

Deliverables and checks

• content/angie/1.md: Agentic Primitives — 1,320 whitespace-delimited words.
• content/angie/2.md: Building an Agent — 1,312 words.
• content/angie/3.md: Agentic Capabilities — 1,368 words.
• content/angie/4.md: Multi-Agent Collaboration — 1,298 words.
• content/angie/5.md: Agent Reliability — 1,375 words.

Every lesson includes learning goals, detailed teaching sections, a worked artifact with reasoned expected behavior, failure conditions, a hands-on exercise with acceptance criteria and rubric, three questions with answers, and focused primary-source links near claims. Markdown uses supported headings, paragraphs, flat lists, inline formatting, links, and fenced code. No source table, raw HTML, image, nested list, or footnote is required.

Only the six assigned files were written by this background author. No commit, push, merge, deployment, live support operation, or secret access was performed.

Kent C. Dodds

Research performed for five independently written lessons using the exact titles and concepts in data/course.json. The input JSON was read, not edited. These lessons do not claim access to MEGA's paid lesson bodies. QueueCare, its specifications, policies, examples, exercises, rubrics, and hypothetical numbers are original teaching material. Fenced examples are illustrative, API-neutral artifacts, not tested runnable code. No transcript was inferred from a video title or thumbnail.

Primary sources used

• You Own the Outcome: Product Engineering in the Age of AI, Kent C. Dodds, September 24, 2026. Public article body read, including output/outcome, shared understanding, options, and Kody example. Supports brief attributed claims in lessons 1 and 2; the remaining teaching is original application.
• Introducing Kody: Your Personal Software Factory, Kent C. Dodds, September 8, 2026. Direct browser-tool opens failed twice. A site-restricted search returned the indexed first-party article body, which was read. Supports the limited descriptions of durable packages and author-described secret isolation in lessons 1 and 3. Product and security assertions are attributed descriptions, not independently audited guarantees. No unviewed video claims used.
• Testing Implementation Details, Kent C. Dodds. Read the article's discussion of refactor brittleness, missed broken interactions, and tests that use public interfaces. Supports lessons 2 and 4. Its accordion implementation and prose were not copied.
• The Merits of Mocking, Kent C. Dodds. Read the discussion of practicality versus lost integration confidence. Supports lesson 3. QueueCare's crash and retry analysis is an original illustrative exercise, not a claim that this source demonstrates those mechanisms.
• Write tests. Not too many. Mostly integration., Kent C. Dodds. Article body read on confidence, testing cost, integration, and coverage. Supports lesson 4. The article title repeats a phrase the article credits to Guillermo Rauch; no claim of Kent originating that phrase is made. No percentage coverage target is taught.
• Should I write a test or fix a bug?, Kent C. Dodds. Read the contrasting product contexts and testing priorities. Supports lesson 4's context-dependent gate selection.
• Testing Library Guiding Principles, official project documentation. Read all principles on DOM and user-like interaction. Supports lesson 4, without presenting semantic queries as a complete accessibility audit.
• Testing Library About Queries, official project documentation. Read query types, accessible role/name priority, labels, and retry distinctions. Supports lesson 4. No version-specific executable snippet is promised.
• Foundations, feedback, and agents — Dillon Mulroy on product at Cloudflare, Chats with Kent, April 1, 2026. Read published first-party transcript sections around 17:34–30:45 and 42:54–47:45, including support relationships, metrics, small scope, and monitoring. Supports lesson 5's single attributed summary. Distinguishes guest comments from Kent's responses; no media playback was performed.
• Pragmatic AI: Kent C. Dodds on MCP and How Product Engineering Might Outlast Coding. Read official show notes and followed its Transcript link.
• Official Pragmatic AI interview transcript, April 22, 2026. Read the review/risk discussion, roughly source lines 289–448, and product-engineering discussion around lines 528–748. Supports lesson 4's short contextual account of review depth. Does not treat personal workflow as a universal release policy.
• OpenTelemetry Signals, official documentation. Read main content at source lines 845–881 defining signals, traces, metrics, and logs. Supports lesson 5's telemetry distinctions. No SDK setup or vendor compatibility claim is made.

Mapping to supplied concepts

• Lesson 1, Mechanics Behind Product Engineering: shift through changing bottlenecks; ownership through the complete booking journey; judgment through competing options; systems thinking through causal links and displaced costs; leverage through reusable capabilities.
• Lesson 2, Translating Intent into Agent Work: intent, context, requirements, specifications, decomposition, and evaluation are treated as distinct stages connected by one complete cancellation contract.
• Lesson 3, AI-Native Product Development: exploration resolves decisions; planning follows uncertainty; execution uses stable boundaries; feedback includes users and checks; engineering examines crashes and integration guarantees; delivery includes activation and recovery.
• Lesson 4, Verifying Product Outcomes: acceptance promises, review, verification, the explicitly original system-diff framework, and consequence-based quality gates form a worked acceptance decision.
• Lesson 5, Products That Learn After Release: signals, observation, synthesis, decisions, and feedback loops connect a hypothetical pilot review to follow-up action.

Seed access limitations and rejected evidence

• MEGA public landing page was accessible. Used only to check public course context; no paid lesson body was available or reconstructed. Exact titles and supplied concepts come from the preserved repository input.
• Kent's supplied YouTube channel returned only a three-line page through the browser tool, without usable video or transcript content. No YouTube viewing is claimed. Published first-party podcast transcripts provide accessible media evidence instead.
• Seed X post 2109860735702425787: browser open returned an internal error; content not used.
• Seed X post 2112046788395880454: browser open returned an internal error; content not used.
• Seed X post 2110835899596681368: browser open returned an internal error; content not used.
• Seed X post 2110815577170378890: browser open returned an internal error; content not used.
• Seed X post 2110027468161921047: browser open returned an internal error; content not used.
• A search result for a Jack Ryan feedback episode redirected on opening to an unrelated seniority episode. This mismatch was rejected as lesson evidence.
• A hybrid-search blog seed discovered during research failed direct opening and was not used in the lessons.
• Search snippets and site lists were used for discovery, except the explicitly recorded indexed first-party Kody article body. No secondary recap supplied lesson claims.

Authorship and checks

All five lessons contain reader orientation, learning goals, detailed teaching, a worked artifact with expected results, failure modes, a concrete exercise with acceptance criteria and a reasoned rubric, three answered questions, adjacent descriptive links, and a sources footer. Source-derived summaries are deliberately brief and spread across primary pages; the bulk is original teaching and original QueueCare examples. No substantive source passage was quoted. Word counts use whitespace-separated Markdown tokens, including headings, links, and illustrative artifacts.

John Lindquist

Research completed October 19, 2026. These notes support five independently written lessons, not the content of the paid MEGA course. The exact lesson titles and topic lists come from the repository's data/course.json. All Doc Steward artifacts, pseudocode, exercises, acceptance criteria, and worked outcomes are original teaching examples. Pseudocode was not executed and does not imply an actual Jev SDK, mdflow configuration schema, installed hook, or deployed service.

Primary material actually read

• Script Kit project site. Read the product description and launch/create/run workflow. Supports lesson 1's concrete automation example. Also opened John's kit repository, whose public README identifies the project and build workflow; no implementation or recovery guarantees inferred from that README.
• Automatically Switch Node Versions Based on Your Package.json and Precmd in ZSH. Read the full short first-person article and code. Supports lesson 2's runtime-mismatch example. Historical Node versions and substring parsing are not recommended as current practice.
• Generate Markdown Links From Your Selected Text and Chrome's Current Url Using Alfred. Read the full short first-person article describing selection, URL capture, and clipboard output. Supports lesson 4's pipeline decomposition. Embedded video content was not available in the extracted page; no claim of watching it.
• Map Alfred Snippets to a Specific Application. Read the first-person article and example mapping frontmost application identifiers to snippet text. Supports lesson 4's context-sensitive routing example.
• What Should I Teach in a Video Tutorial?. Read all substantive sections: target skill, context, instructions, demonstration, and recap. Supports lesson 5's outcome-based feedback framing. Evaluation design itself is our synthesis.
• johnlindquist/claude-hooks. Read public README overview, requirements, generated layout, and sample handlers. Supports lesson 4's typed handler example and Bun requirement. Did not install the package or verify compatibility against a local Claude Code runtime. Current official runtime documentation is used for lifecycle semantics.
• johnlindquist/mdflow. Read README sections on Markdown execution, engine selection, environment/context isolation, and proposal-first evolution. Supports lesson 2's portability discussion. No claim that our sample manifest is mdflow syntax.
• mdflow SECURITY.md. Read execution-model section. Supports lesson 2's distinction between dry-run inspection, context isolation, and host sandboxing.
• mdflow docs/evolve.md. Read change-with-proof overview, quick start, outcome definitions, durable evidence, policy modes, and verification receipt introduction. Supports lesson 5's proposal/application separation and fix-versus-regression-safe distinction. Did not execute its commands or paid evaluations.

Primary documentation used to support mechanisms

• Temporal workflow execution. Read execution definition, reliability, replay, event history, commands, statuses, and state transitions. Supports lesson 1. Our small state machine borrows concepts and does not claim Temporal-equivalent guarantees.
• TypeSafe introduction. Read the public typed-question contract, Choice/Score/Noul distinction, independent same-state questions, and atomic-question guidance. Supports lesson 1. Read TypeSafe's main site for discovery only; do not repeat marketing performance or zero-hallucination claims as verified evidence.
• TypeSafe confidence. Read probability/confidence distinction, risk-sensitive threshold guidance, Noul difference, and Choice/Score formula sections. Supports lessons 3 and 5. No invented return schema or API call appears in the lessons. A confidence value is not represented as a guarantee of correctness or action authorization.
• Claude Code hooks reference. Read lifecycle, event timing, matchers, configuration locations, and pre-tool decision example. Supports lesson 4. Current source redirects from the older docs.anthropic.com URL. Lesson limits its product-specific claims to event timing and typed adapter boundaries.
• GitHub webhook best practices. Read all substantive sections: scope, secrets, HTTPS, prompt response/queue, event/action validation, missed delivery, delivery identifier. Supports lesson 4. Durable inbox and semantic deduplication design is original synthesis, not claimed GitHub infrastructure behavior.
• Building effective agents. Read routing, parallelization, orchestrator-workers, evaluator-optimizer, and agent-loop sections. Supports lesson 3's decomposition. Article is historical and contains a current tooling caveat; no framework recommendation is inferred.
• Demystifying evals for AI agents. Read evaluation structure, grader types, capability/regression distinction, balanced cases, isolation, grading outcomes rather than prescribed tool paths, and human calibration. Supports lesson 5. No reported benchmarks from the article are reused.

Instructor media: accessible descriptions, not watched media

• Advanced Claude Code techniques: context loading, mermaid diagrams, stop hooks, and more. Read host-published show notes, topic list, and chapter links for John's interview. Supports only the statement that context loading and documentation are listed topics in lesson 2. The page shows a transcript heading but no substantive transcript in available extraction.
• Linked YouTube interview. Opened the public video page; extraction exposed only minimal page material, not the video or transcript. No claim of watching or hearing it.
• Jev: 8 real use cases, How I AI / Apple Podcasts. Read public episode title, show description, listed use cases, and chapter labels, including routing and multi-agent coordination. Supports lesson 3's explicitly limited media-topic statement. Did not listen to the audio or obtain a transcript. Product contract claims rely on TypeSafe documentation instead.

Seed-source access limitations and exclusions

• MEGA rethinking AI with Jev: browser tool returned inaccessible. No content claims based on it.
• jevhooks plugin/hooks path: browser fetch returned cache miss. No claims based on its code. John's public claude-hooks repository was read instead.
• John X post 2110781910272360924, post 2110562239518519754, and post 2110145728152826147: each open returned an internal/inaccessible result with no post text. No quoted or paraphrased post claims used.
• self.md profile/summary: read for discovery of John's repositories only. It is a secondary recap, so its payload examples, biography, usage counts, and API claims are not treated as evidence. Followed the links to firsthand repositories instead.
• Paid course bodies, paid recordings, and private repositories were not accessed or reproduced. No lesson claims to be John's own script, exact spoken teaching, or an official MEGA companion.

Lesson mapping and interpretation boundaries

• Lesson 1: Script Kit, Temporal execution, TypeSafe introduction. Original emphasis: scoped autonomy, state machine, effect reconciliation, meaningful progress observation.
• Lesson 2: mdflow README and security policy, John's Node-selection post, limited interview show notes. Original emphasis: manifest layering, capability adapters, preflight, provenance, equivalent behavior across environments.
• Lesson 3: public Jev episode description, TypeSafe confidence, Anthropic agent-pattern article. Original emphasis: artifact ownership, dependency graph, leases/generations, bounded supervision.
• Lesson 4: John's two Alfred articles and claude-hooks README, official runtime hook documentation, GitHub webhook guidance. Original emphasis: durable inbox, semantic task identity, backpressure, pre-effect enforcement.
• Lesson 5: John's teaching post and mdflow evolution protocol, Anthropic eval article, TypeSafe confidence. Original emphasis: causal feedback triage, reviewed fixtures, workflow distillation, versioned promotion and rollback.

All source-dependent summaries are short and distributed among primary sources. Most text develops original examples, tradeoffs, exercises, and explanations. Source links appear beside the claims they support and in each lesson footer. No direct quotations or paid lesson excerpts are included.