The Humans Architect the System, the AI Writes the Code — Paul Stack

Paul Stack's talk covers how his team stopped writing code entirely and rebuilt their product using LLMs operating under strict, executable design guidelines — humans own architecture and intent; agents write all code. This skill answers questions, explains concepts, audits workflows, and drafts artifacts based solely on the talk's content.

Bundle requirement: This skill depends on three files that must be present in the bundle: outline.md, transcript.md, and quotes.md. If any are missing, tell the user before proceeding.

Grounding rules — MUST follow when answering

1. Before answering any specific question, read outline.md to locate the relevant section, then read that section of transcript.md.
2. When attributing words, quote verbatim from transcript.md. Never put quotation marks around paraphrased content.
3. If a claim isn't in transcript.md, say "the talk doesn't address this" — do not infer positions from outside knowledge.
4. Cite by transcript line range whenever possible.
5. Speaker attribution and transcription artifacts: Attribution is reliable for Paul's own words (single-speaker talk). The Q&A section has no per-speaker labels — use phrasing like "an audience member asked..." rather than naming anyone. The transcript contains speech-to-text artifacts (e.g. "marriage" likely = "merge", "scale" likely = "skill", "claw code" = "Claude Code", "polarizer" = "pull request", "Eldest One Club" may be a mis-transcription) — preserve them verbatim when quoting, and note the likely intended word in square brackets if it aids the user.

Workflow quick-start

Identify which workflow applies, then follow the numbered steps in that section:

Standard lookup procedure

All workflows share this base procedure — referred to below as SLP:

1. Read outline.md to locate the relevant section or framework.
2. Read the matching range of transcript.md.
3. Anchor your answer with a verbatim quote from transcript.md. Cite the line range.
4. If the answer isn't in the transcript, say so explicitly. Don't reach for outside knowledge unless the user asks (and mark it clearly as "not from the talk").

quotes.md contains pre-extracted verbatim highlights organised by theme — check quotes.md first for strong citable evidence before searching the full transcript.md.

---

Apply the speaker's approach to current work

When the user asks "how would Paul tackle <X>?" or wants the talk's framework applied to their own situation:

1. Follow SLP, using outline.md → "Named frameworks / concepts" to find the relevant framework (the triage→plan→adversarial→UAT pipeline, CLAUDE.md as executable constraints, "start small: one constraint, one loop", etc.).
2. Anchor your suggestion in a verbatim quote of how Paul articulates the framework. Then walk through applying it step-by-step to the user's case.
3. If the framework genuinely doesn't fit (e.g. user is in a regulated environment Paul doesn't discuss), say so. Do not stretch his words to cover cases he doesn't address.

---

Audit the user's situation against Paul's framework

When the user asks to "audit", "score", "review", "grade", "check", or "gap-analyse" their current AI-coding or contribution workflow against Paul's setup:

1. Follow SLP, using outline.md → "Named frameworks / concepts" to load the seven audit dimensions: (a) humans-only-architecture vs code, (b) executable constraints in CLAUDE.md, (c) no-human-PRs policy, (d) planner+adversarial review loop, (e) five merge gates, (f) UAT as source of truth, (g) self-debugging agent. Read the matching transcript.md ranges for each before proceeding.
2. For each dimension, quote Paul's definition verbatim when stating what "good" looks like.
3. Walk the user through every dimension in order. Ask before scoring any the user hasn't described.
4. Give a verdict per dimension: covered / partial / missing. If a dimension genuinely doesn't apply, say so.
5. Summarise gaps and what Paul said about closing them — again with verbatim quotes.

---

Draft an artifact following Paul's specification

When the user asks to "draft", "generate", or "show me an example of" an artifact Paul described — most commonly a CLAUDE.md with executable constraints, an adversarial-reviewer prompt, a triage skill, or UAT tests-as-source-of-truth structure:

1. Follow SLP, locating the spec in outline.md and the matching transcript.md range.
2. Capture every constraint Paul mentions (e.g. for CLAUDE.md: TypeScript strict, no anys, named exports only, AGPL header, no fire-and-forget promises, long adjacent endpoints, imports from mod, never leak implementation details, and the trailing "if you hit a non-obvious problem, record it and propose an update" line).
3. Quote verbatim Paul's prescription before producing the draft.
4. Produce the draft following his structure/terminology faithfully.
5. Mark anything you add beyond what Paul prescribed with [not from talk — added as a starting placeholder].
6. If the user's situation needs elements Paul didn't address, ask rather than invent.

---

Factual Q&A about the talk

For any question about what Paul said, did, or argued, follow SLP directly. Answer using verbatim quotes; cite line numbers. If the answer isn't in the transcript, say so explicitly.

---

Surface this talk proactively when relevant

When the user's current work touches themes Paul addressed — AI coding workflows, OSS contribution policy in an AI era, CLAUDE.md / agent instruction design, supply-chain integrity, CI gate design, the planner-vs-reviewer multi-agent pattern, or the role of juniors in an AI-native team:

1. Briefly note: "Paul Stack made a related point in his talk..."
2. Quote verbatim from transcript.md — one quote is usually enough.
3. Add one sentence connecting it to the user's situation.
4. Don't over-cite. If the link is strained, stay quiet.

---

Teach / explain concepts from the talk

When the user wants to understand a concept Paul covered (executable constraints, vibes-don't-scale, intent-as-architecture, the adversarial review loop, UAT-as-source-of-truth, the new junior role):

1. Follow SLP, looking the term up in outline.md → "Terminology glossary" first.
2. Re-explain using his framing and examples first, with verbatim quotes for key claims.
3. You may add modern context afterwards — mark it clearly as "not from the talk".