CurioQuest — Operation Log

[2026-06-29] new | CurioQuest

Files touched: index.md, log.md, AGENTS.md, wiki/overview.md, wiki/requirements.md, wiki/open-questions.md Notes: Wiki structure initialized around existing vault folder.

[2026-06-29] record | Claude Code OAuth text-generation provider

Files touched: index.md, log.md, wiki/dev-log.md, wiki/requirements.md, wiki/open-questions.md, wiki/decisions/2026-06-29-claude-code-oauth-text-generation.md Notes: Captured the local/sample-generation provider pivot, verification, and remaining product-quality build list.

2026-07-03 — Quality loop ran to terminal (Fable 5, autonomous /loop)

All 4 artifacts PASSED at Round 0 on the v5 prompts verbatim (sound_quest story 4.91 / experiment 5.00; wilting_mystery story 4.97 / experiment 4.81). Deterministic gates (procedure-spine, K age-veto, safety pre-gate) all clean, verdicts from code. WINNER prompts = source unchanged — nothing to apply. Caveats: validated on Fable 5 (production is gpt-5 — re-verify there); generator==judge this run. Durable findings for founder: K lesson packet emits four-cup factorial procedures (K-illegal), wilting packet has empty target_standard_code, sound_quest packet activity (band-on-cup) collides with the story plot under D18. Full record: repo docs/workspace/quality-loop/SUMMARY.md.

2026-07-03 (late) — Quality loop RE-RUN valid; winners applied to branch

First loop run this session was invalid (generator==judge, gamed gates, self-scored PASS; spec had a founder-removed Round-0 exit) — quarantined. Re-ran with blind-subagent harness (generator sees only prompt+book_context; judge never generates; Round 0 never terminates). All 4 artifacts PASSED via real diagnose→patch→regenerate cycles. Winners APPLIED to apps/website/src/lib/bookGenerationPrompts.ts on branch feat/quality-loop-winner-prompts (uncommitted): story P2 plausibility-law + P3 deterministic story-scan; activity word-scan + homograph-trap + letter-scan. P1 hard word ceiling applied then REVERTED (collides with locked D8 "word count is guidance not a gate"; book length is founder taste). tsc clean, 14/14 tests pass. Open: item C (sentence-length grade-fit gate) awaits founder block-vs-warn + thresholds. Nothing committed/pushed. Full record: repo docs/workspace/quality-loop/SUMMARY.md.

2026-07-07 — 🔴 PIVOT: pipeline discarded, redesigning as editor-in-chief craft studio

Founder scored the 2nd produced book (Water Cycle) -1/10 and rejected the whole automated-pipeline approach: incoherent dialogue, characters flickering, Professor T-Rex absent, ugly pdf-lib text blocks, no visual flow. Diagnosis (mine, owned): automated end-to-end before ever producing one good page; story built by a disconnected story→page_plan→dialogue relay; pdf-lib is the wrong layout tool; images generated independently with no storyboard; vision-QC was theater (passed a book with no T-Rex). Reset via superpowers:brainstorming — KDP Phase 1 reframed as making 7 excellent FIXED books (a craft problem, not the personalization pipeline). Locked: editor-in-chief studio (baseline-before-automation); writing = founder premise+setting → AI coherent manuscript → founder edits; teaching model C (guide teaches warmly + kids do + small aha, reversing D17); art = storyboard-first (proposed). Also this session before the pivot: OpenAI fully removed → Gemini for images (nano-banana gemini-3.1-flash-image) + vision-QC; operator console + KDP landing live locally; KDP research pack (YT + Skool + SPARK analysis) written; monorepo restructure built in worktree (unmerged). Brainstorm stopped mid-Question-4 (art approach + style). Files: repo ./SESSION.md (full), docs/Project-Objective.md (source of truth), docs/research/kdp/**. Next: finish brainstorm → design doc → writing-plans → build the studio → produce ONE great book as baseline.

2026-07-07 (cont'd) — Design + plan approved; Character Bible LOCKED (Phase 0 done)

Continued from the pivot logged above. superpowers:brainstorming ran to completion: founder answered 4 forked questions (philosophy A editor-in-chief; writing A+B mix — founder premise + AI manuscript; teaching model C — guide teaches + kids do; art storyboard-first) then picked art style (Style 4, classic rich gouache, from a 4-way side-by-side sample of the hilltop premise scene) and layout style (A, classic storybook — art and text each get their own clean space, no bubbles). Design doc written + approved (docs/superpowers/specs/2026-07-07-curioquest-studio-design.md), then the 6-phase implementation plan (docs/superpowers/plans/2026-07-07-curioquest-studio.md, subagent-driven execution chosen). Model routing clarified per founder ask: Fable 5 pinned to ONLY the manuscript/story-writer role (frontier reasoning belongs where craft lives); Sonnet for code-building subagents; Opus (this session) for orchestration and editorial judgment.

Phase 0 executed and LOCKED same session: generated gouache cast-sheet turnarounds for all 4 characters via a new Workflow/scripts/studio/gen-cast-sheets.mjs (Gemini gemini-3.1-flash-image). Went through several founder-driven correction rounds — Ronav's look (based on a real photo of the founder's son) required two fixes: a first "mohawk" attempt wrongly shaved the head bald on the sides (corrected by re-deriving from the actual reference photo: full messy hair, no shaved sides, and his real gap-tooth smile — missing front teeth — which had been missed entirely); Zara/Sam were caught as "the same face in different clothes" and rebuilt with genuinely different skin tone, face shape, nose, and eyes. Also locked: ages 7-8 for all three kids, no 4th character, Ronav's wardrobe (blue jeans, white ROAR T-Rex tee, plain blue sneakers with the Nike swoosh deliberately dropped for KDP trademark safety, and a toy T-Rex as his fixed prop instead of a magnifying glass), Professor T-Rex's white coat + pocket-watch (swapped from an earlier tweed-vest draft). Committed 81799a7 and pushed. Fix-and-learn logged: image-reference regeneration must always re-derive from the actual source image, never from a prior text description of it — the same failure class as the original pipeline's core defect (skipping grounding in the real reference), caught fast here because of the per-artifact founder gate.

Next: Phase 1, the manuscript — Fable 5 writes the Water Cycle story from the founder's hilltop premise; founder edits hard (3-5 rounds expected) until genuinely good. Full detail: repo SESSION.md, docs/superpowers/plans/2026-07-07-curioquest-studio.md.