Data Scientist · L0011 · 2026-07-17

Slice data-scientist/0011 · Level 0011 (AI-Assisted Development) · Apprentice tier · Engine verdict ⚠️ warn · Walked 2026-07-17

🔗 Perfection run · 🏠 Perfection dashboard · 📄 Raw report · 🕘 Change history

🎯 Session Summary

I walked the full Data Scientist · Level 0011 (AI-Assisted Development, 🌱 Apprentice) slice — all 4 main quests the planner selected, in plan order — as a learner would, using the workflow-sealed execute-mode evidence plus a close read of each quest source. Headline: warn — the slice is mostly usable, but it is not a clean linear path and it contains one hard, blocking failure.

Three quests hold up when their commands are actually run (Summon the Golem passes cleanly at 88; Hidden Gem and Prompt Crystal warn at 68/64 with fixable defects). The one that breaks a learner is The PRD Codex (43, fail): its entire hands-on core hangs on docker compose build prd-machine, which the engine verified fails with “no configuration file provided” in any repo that doesn’t already ship the PRD Machine tooling — and the quest never shows how to obtain that tooling. A maintainer’s most valuable next action is to make PRD Codex self-contained (or gate it behind the repo it assumes); everything else in the slice is polish.

Note on cohesion: these four quests share the 0011 level but belong to four different quest series/lines (The Ouroboros Loop, Documentation Mastery, Web Publishing, AI Development Mastery). This is a themed bundle, not a single authored chain — an important framing for the continuity findings below.

🗺️ The Journey

Plan order (dependency-sorted by the planner):

1. ✅ Summon the Golem: An AI Agent Joins the Loop — 88/100 (pass) · Drive headless Claude Code from GitHub Actions under a bounded role. Technically tight: every CLI flag, env var, and Action version checked out; the gating/no-op bash ran exactly as claimed. Only gaps are a missing env-wiring reference and no answer key for the knowledge checks.
2. ❌ The PRD Codex: Master Product Reality Distillation — 43/100 (fail) · The pivotal docker compose build prd-machine was verified to fail with “no configuration file provided” — the quest never provides or points to the PRD Machine tooling, so Chapters 2–3 and all four Challenges are unrunnable from the document alone. Blocking.
3. ⚠️ Hidden Gem: Publish AI Chats on GitHub Pages — 68/100 (warn) · Jekyll config, Gemfile, AI-chat post, and bundle exec jekyll serve genuinely work; a Linux bundle install permission gotcha (no vendor-path/.gitignore guidance) plus a wrong theme URL and a suspect extension ID drag accuracy down.
4. ⚠️ Forging the Prompt Crystal: VS Code Copilot Mastery — 64/100 (warn) · Cross-platform scaffolding and both Mermaid diagrams run; but {​% raw %​} Jekyll build artifacts leak into the .prompt.md templates, several prompt blocks are mislabeled javascript and error as JS, and the Windows/Cloud setup silently overwrites an existing copilot-instructions.md.

🔬 Evidence

All results are from execute mode (walk-evidence.json, sealed by the workflow — I consumed it as-is and did not re-run the engine). Snippet counts are ran / passed / failed / skipped / reasoned. Dimension scores are on the 1–5 rubric (commands_work, content_accuracy, completeness, clarity, structure, safety).

1. Summon the Golem — 88 ✅ (5/5 passed, 1 skipped, 2 reasoned)

Dimensions: commands_work 4 · content_accuracy 5 · completeness 4 · clarity 4 · structure 5 · safety 5

• passed — Composite action.yml and the workflow step YAML fragments parse cleanly with pyyaml (no syntax errors).
• passed — The no-auth guard if [ -z "${CLAUDE_CODE_OAUTH_TOKEN:-}" ]; then … exit 0; fi was extracted and run with the var unset: it printed the ::warning:: line and exited 0, exactly as the quest claims (“No auth = exit 0 with a warning”).
• passed — The gate logic (go=false; if [ "$ENABLED" = "true" ] && [ -n "$OAUTH" ]; then go=true; fi) tested across all 4 ENABLED/OAUTH combinations — arms only when both true, matching stated intent.
• passed — Mermaid quest-network diagram compiled to valid SVG.
• reasoned — claude -p … --append-system-prompt … --allowedTools … --permission-mode acceptEdits, npm install -g @anthropic-ai/claude-code, claude setup-token are all valid/current commands (verified against docs, not executed end-to-end — a full CI run needs a real repo + secret).
• skipped — .claude/agents/potion-scribe.md role file (markdown, nothing to run).

2. The PRD Codex — 43 ❌ (13 ran: 5 passed, 8 failed, 2 skipped, 4 reasoned)

Dimensions: commands_work 1 · content_accuracy 2 · completeness 2 · clarity 3 · structure 4 · safety 4

• failed — Cloud Realms path cd /workspaces/it-journey; docker compose build prd-machine; … — build fails: “no configuration file provided: not found”. The same docker compose build prd-machine recurs at source lines 182, 208, 234, 249 across the macOS / Linux / Cloud paths; none can succeed without pre-existing tooling the quest never supplies.
• failed — All downstream docker compose run --rm prd-machine …/prd-machine sync (incl. --days 7, --output /tmp/custom-prd.md), status, conflicts, and the “Quick Win Checkpoint” sequence — every one fails because the container never built.
• failed — macOS and Linux path bash blocks (docker --version; … cd /path/to/your/repository; cd ~/github/it-journey; …) error as written (contradictory back-to-back cds, unbuildable target).
• failed — “Best Practices for Commit Signals” block (~line 384) is fenced bash but is commit-message prose; run as bash it throws “syntax error near unexpected token ‘auth’“.
• passed — Static content is sound: Signal-Rich Frontmatter YAML, features/features.yml, and all three Mermaid diagrams (Signal Source Architecture, Quest Network Position, Implementation Flow sequence) parse/render correctly.
• reasoned — The “Expected Output” transcripts (two mislabeled sql, actually log text) are repo-specific sample output, not literal expected results.

3. Hidden Gem: GitHub Pages — 68 ⚠️ (4 ran: 3 passed, 1 failed, 3 skipped, 1 reasoned)

Dimensions: commands_work 3 · content_accuracy 3 · completeness 3 · clarity 4 · structure 4 · safety 5

• passed — _config.yml (minima), the captured AI-chat post _posts/2025-11-14-ai-quest-chat.md, and bundle exec jekyll serve build/serve — the hands-on core genuinely works.
• passed — _config.yml with plugins: [jekyll-feed, jekyll-sitemap] is valid.
• failed — Gemfile + plain bundle install: fails with a PermissionError on distro/system Ruby; the quest offers no bundle config set --local path vendor/bundle fallback and no .gitignore for vendor/bundle, _site/, .bundle/.
• skipped — OS-specific install blocks (brew / winget / apt + code --install-extension …) — environment-dependent, not run.
• reasoned — AI Conversation Capture markdown template is well-formed.

4. Forging the Prompt Crystal — 64 ⚠️ (10 ran: 5 passed, 5 failed, 4 skipped, 21 reasoned)

Dimensions: commands_work 3 · content_accuracy 3 · completeness 4 · clarity 3 · structure 4 · safety 3

• passed — Setup scaffolding runs: mkdir -p .github/prompts && touch .github/copilot-instructions.md (Linux), the Cloud echo '# …' > … variant, and the Windows PowerShell code --install-extension … / Select-String block. Both Mermaid diagrams (Quest Network, Implementation Flow) render.
• failed — Five ```javascript-fenced blocks (source lines ~194, 205, 581, 948, 953) are plain-English prompt text and throw SyntaxErrors when linted/run as JS.
• reasoned — The RCTF/CoT/few-shot templates and the .prompt.md template bodies (lines ~683–839) are prose/markdown: static-checked. These carry the two high-severity defects below (leaked {​% raw %​} tags; non-native prompt-file syntax).
• skipped — macOS/Linux code --install-extension and code --list-extensions | grep blocks (no code CLI in sandbox / environment-dependent).
• safety = 3 — flagged because the Windows/Cloud setup silently truncates an existing copilot-instructions.md via -Force / > with no warning.

🐞 Issues Found

Severity · quest · where · observed · suggested fix. (Every item below is grounded in a command the engine actually ran or an exact line in the quest source.)

• HIGH · PRD Codex · Chapters 2–3, docker compose build prd-machine (lines 182, 208, 234, 249) — Verified failure “no configuration file provided: not found” in a fresh repo; the PRD Machine tooling is never provided. → Add an explicit bootstrap step (git clone URL, or a minimal docker-compose.yml + Dockerfile + scripts/prd-machine snippet) before the build, so a learner starting fresh can actually reach sync/status/conflicts. This alone is the difference between a fail and a pass for the quest.
• HIGH · PRD Codex · “Best Practices for Commit Signals” (~line 384) — bash fence around commit-message prose; running it throws “syntax error near unexpected token ‘auth’“. → Re-fence as text or git-commit.
• HIGH · Prompt Crystal · Chapter 4 .github/prompts/ templates — Leaked {​% raw %​}…{​% endraw %​} Jekyll escape tags (12 occurrences) in the code-review, debug-assistant, and test-generator bodies; copied verbatim they pollute a learner’s real .prompt.md. → Strip the raw tags from the rendered template content.
• HIGH · Prompt Crystal · Chapter 4 prompt-file syntax — The shown name/version/inputs/{​{ inputs.x }​} convention is not VS Code Copilot’s native prompt-file spec (mode/tools/description frontmatter, ${input:var}). → Align with the native spec, or explicitly label it an IT-Journey doc convention Copilot won’t resolve.
• MEDIUM · Hidden Gem · Chapter 3 bundle install — Verified PermissionError on system Ruby; no vendor-path fallback, no .gitignore before the Step 6 commit (risks committing tens of thousands of dependency files). → Add bundle config set --local path vendor/bundle guidance + a .gitignore (vendor/bundle, _site/, .bundle/) and recommend rbenv/rvm.
• MEDIUM · Hidden Gem · line 148 theme link — themes.jekyllrc.org contradicts the quest’s own Resource Codex (which correctly links jekyllthemes.io). → Fix to the real site / jekyllrb.com/docs/themes.
• MEDIUM · Hidden Gem · VS Code extension IDs (lines 47, 62) — “GitHub Pull Requests and Issues” ID looks wrong (ms-vscode.…); it lives under the GitHub publisher namespace. → Verify and correct.
• MEDIUM · Prompt Crystal · Windows/Cloud setup file-overwrite — New-Item … -Force -Path '.github\copilot-instructions.md' and echo … > …/copilot-instructions.md silently truncate an existing populated file (verified). → Add a warning / guard for “your current project.”
• MEDIUM · PRD Codex · Challenge 4 — requires reviewing .github/workflows/prd-sync.yml that is never shown/linked. → Include the workflow (or a representative excerpt).
• LOW · Golem · env wiring — The composite action reads CLAUDE_CODE_OAUTH_TOKEN from job env but this chapter never shows the env:/secrets wiring. → Add a one-line reference to where the token is injected.
• LOW · Golem · knowledge checks — Open-ended questions have no answer key for self-paced learners. → Add collapsible answers.
• **LOW · Prompt Crystal · 5 javascript blocks (lines 194, 205, 581, 948, 953)** — plain prompt text; SyntaxError if run as JS. → Re-tagtext/markdown```.
• LOW · Prompt Crystal · RCTF vs. examples — 4-letter “RCTF” but nearly every worked example adds a 5th [CONSTRAINTS] section. → Rename (e.g. RCTFC) or note Constraints as an optional 5th block.
• LOW · PRD Codex · code-fence languages — two sql blocks (lines 248, 300) are plain log lines; a macOS block has contradictory back-to-back cds. → Re-tag text and drop the duplicate cd.

🔗 Chain Continuity

This slice is a level bundle, not a single authored chain. The four quests carry four different quest_series: The Ouroboros Loop (Golem), Documentation Mastery (PRD Codex), Web Publishing (Hidden Gem), AI Development Mastery (Prompt Crystal). A Data Scientist arriving at 0011 is not walking one story — they’re picking among parallel AI-assisted-development tracks. Judged that way, the theme coheres (everything is about wielding AI tooling), but the narrative does not.

Prerequisite reality for a learner walking in plan order:

• Golem assumes “your potion-book repository with the Warden’s Gate in place” (Chapters I–II of the Ouroboros Loop, which live at level 0101 — outside this slice) and a claude setup-token OAuth token. Its quest_dependencies.required_quests is empty, so nothing in-slice or in-frontmatter surfaces that gate. A learner who starts here cold has no gated loop to attach the golem to — the quest teaches the agent well in isolation, but its assumed repo state is unmet by the slice.
• PRD Codex unlocks_quests explicitly lists prompt-crystal and github-pages — the only real intra-slice dependency edge, and the planner honored it by ordering PRD before those two. But PRD is exactly the quest that fails: its Docker core is unrunnable, so the “unlock” is hollow — a learner following the intended order hits a wall on quest 2 and cannot legitimately complete it before moving on. The good news: Hidden Gem and Prompt Crystal don’t actually consume any artifact PRD produces, so the wall is a motivation/credential break, not a hard technical block on quests 3–4.
• Hidden Gem requires /quests/0000/hello-noob/ (an earlier level — reasonable and met by normal progression) and is the gentlest, most self-contained entry (🟢 Easy) — arguably the best first quest for a newcomer to this level, ahead of the Medium ones.
• Prompt Crystal requires /quests/0010/prompt-engineering-mastery/ (prior level). It stands alone fine; its defects are content-quality, not continuity.

Ordering observation for a maintainer: as a learning path for a Data Scientist, Hidden Gem (Easy, self-contained) → Prompt Crystal (prompt skills) → Golem (agent in CI, but flag the 0101 Warden’s-Gate prerequisite) is a smoother ramp than the current dependency-sort, and PRD Codex should be treated as not yet walkable until its tooling bootstrap is added. No quest in the slice silently depends on an artifact produced by another that I could not otherwise obtain — the friction is (a) the unmet Warden’s-Gate repo state Golem assumes, and (b) PRD’s missing tooling.

🧠 Reasoning & Method

• Mode: execute. I did not run the engine myself — per the skill and the CI contract, the workflow pre-computed and sealed walk-evidence.json / walk-evidence.md (the engine’s child claude processes can’t authenticate from an agent’s Bash tool). I consumed the sealed evidence verbatim and did not edit, regenerate, or hand-write any number. walk-plan.json and walk-evidence.* were left untouched.
• What was actually run (in the sandbox, by the sealed engine): 32 recorded commands across the four quests, of which the engine ran ~32 and reasoned statically about the rest — Golem 5/5 passed; PRD Codex 5 passed / 8 failed; Hidden Gem 3 passed / 1 failed; Prompt Crystal 5 passed / 5 failed. Every passed/failed above traces to one of those runs; every reasoned item is labeled as static-only.
• What I reasoned about (me, statically): the linked-journey continuity — series cohesion, prerequisite satisfaction, and ordering — by reading each quest’s frontmatter and body in plan order. Those are the §🔗 findings and are explicitly reasoning, not fresh command evidence.
• Coverage / limits: Nothing was capped by me (the planner’s window covers the full 4-quest level; truncated: false). OS-specific installs and code-CLI steps were skipped by the engine as environment-dependent, so cross-platform install accuracy is only partially witnessed. PRD Codex’s Docker path is genuinely broken in a clean repo, so its downstream chapters are unverifiable by design of the defect — that is itself the finding, not a coverage gap on my side. No network-dependent or destructive steps were run.
• Confidence: High on the per-quest verdicts (they come from real sandbox runs with quoted failure output). High on the PRD blocking-fail and the two high-severity Prompt Crystal defects. Medium on the extension-ID / theme-URL slips (asserted by the engine’s content check, not independently re-verified here). The overall slice verdict is warn: three of four quests are usable today, but the slice contains one hard fail that a content pass should fix before this level is called “perfect.”

Machine evidence header (verbatim from walk-evidence.md): “4 quests evaluated · ✅ 1 pass · ⚠️ 2 warn · ❌ 1 fail · avg 65.8% · ~$6.4993”. One slice, one report — findings only; a content pass (content-curator / human) acts on the issues above.