Store Page FAQ LLM Draft Pipeline - Human Diff Gates Before Q3 2026 Partner Review

You asked a model to "write our Steam FAQ." It invented co-op cross-play, cited a 2024 roadmap you never shipped, and used the word unlimited next to procedural content. Q3 2026 partner review compares FAQ bullets to demo builds and AI disclosure rows—ungoverned LLM paste is a faster fail than a crash log.

This AI Integration / Workflow article defines a store-page FAQ draft pipeline: frozen source bullets, LLM-assisted wording only inside guardrails, mandatory human diff gates, and exports into release-evidence/ai-disclosure/faq-pipeline/ before upload. It complements—not replaces—the 14-day prompt registry freeze and patch-notes human loop.

Who this is for and what you get

Time: ~2 hours to stand up pipeline; 20–40 minutes per FAQ refresh after templates exist.
Prerequisites: List of honest feature bullets, human reviewer (cannot be the same person who clicked "generate" without a second sign-off on high-risk rows).

Why this matters now (May 2026)

1. Partner dimension 7 — Mock audit tabletop scores FAQ against demo and disclosure annex.
2. Fest store truth — Wishlist truth audit treats FAQ as binding copy.
3. LLM speed trap — Teams use models for October Next Fest copy refresh; speed without diff gates recreates 2024 "early access forever" embarrassment.
4. Prompt registry gap — Registry freeze governs runtime prompts; store FAQ is marketing surface—needs its own gate.
5. Evidence maturity — BUILD_RECEIPT culture expects auditable promotion; FAQs should match.

Direct answer: No LLM output reaches Steamworks until it passes Gate D (human diff) with signed faq-promotion-log.md.

Pipeline overview (five gates)

Never reorder: A → B → C → D → E.

Roles and separation of duties

Micro-teams overlap roles—still never same person Gate C + sole Gate D on AI/price rows.

Gate A — Freeze source bullets (human-only)

Goal: Model may rephrase; it may not invent.

Create faq-source-bullets-v1.md with only facts a human will defend:

# FAQ source bullets (frozen 2026-05-18)
- Platforms: Windows 10/11 (Steam)
- Players: single-player; no co-op in demo
- AI: NPC barks use cloud LLM; names in ai-disclosure annex
- Demo ends: after chapter 1 (~45 min)
- Price: $19.99 USD base (see regional sheet)
- NOT in demo: crafting, multiplayer

Rules:

• One bullet = one observable fact or explicit negation.
• No marketing adjectives ("best," "ultimate") in source file.
• Version with date; bump v2 only when build changes scope.

Pass: Engineer agrees every bullet is testable on fest-rc branch.

Gate B — Lock prompt template

Store FAQ template example (adapt; keep banned-claim list):

You are drafting Steam store FAQ answers for a PC game.
RULES:
- Use ONLY facts from SOURCE_BULLETS below.
- If SOURCE_BULLETS lacks an answer, output [NEEDS_HUMAN].
- Ban words: unlimited, guaranteed, always, never (unless in source).
- Max 80 words per answer.
- Plain language; no hype.
SOURCE_BULLETS:
<<<paste frozen file>>>
QUESTIONS:
<<<numbered list>>>

Save template hash in faq-prompt-template.lock (sha256 of file). Any template change resets Gate D for all answers.

Pair banned-claim list with Steam/Play disclosure checklist.

Gate C — LLM draft (machine step)

Temperature and model discipline

Do not chain multiple models in one Gate C pass without re-running Gate D—tone and claim drift multiply.

Retrieval-augmented drafts (advanced)

If you attach design docs to retrieval:

1. Chunk docs with stable ids (GDD-sec-12).
2. Require model to cite chunk ids per sentence.
3. Human verifies chunk still true on current build.

Without citation discipline, retrieval only speeds up hallucinations.

Allowed uses:

• Rephrase for readability
• Generate question list from bullets (human trims)
• Produce diff-friendly markdown output

Forbidden:

• Auto-publish to Steamworks
• Web search or repo access unless you implement retrieval with citations humans verify
• Single-shot "write entire FAQ" without per-question review

Output path: faq-drafts/YYYYMMDD-model-draft.md — never overwrite source bullets.

Log model id + date in draft header for partner questions.

Gate D — Human diff gates (mandatory)

Goal: Every sentence traceable or rejected.

Workflow per question:

1. Open draft side-by-side with faq-source-bullets.
2. Highlight claims not in bullets → delete or add bullet first.
3. Run diff (git, Meld, VS Code) against last published FAQ if refreshing.
4. Second human signs high-risk rows: AI, price, platform, multiplayer, demo scope.
5. Append row to faq-promotion-log.md:

| date | question_id | reviewer | diff_ref | pass |
|------|-------------|----------|----------|------|
| 2026-05-18 | Q03_ai | lead | git abc123 | yes |

High-risk triggers (always two humans):

• Any AI / ML / procedural wording
• Price, DLC, or refund language
• Multiplayer / cross-play
• "Full game" vs demo boundaries
• Accessibility or compliance promises

Pass: Zero [NEEDS_HUMAN] left unresolved; promotion log complete.

Gate E — Store parity scan

Cross-check approved FAQ against:

Use truth audit Day 3 questions as scan script.

Pass: Producer signs faq-parity-scan-2026.md with build id from BUILD_RECEIPT.

Folder layout

release-evidence/ai-disclosure/faq-pipeline/
  faq-source-bullets-v1.md
  faq-prompt-template.lock
  faq-drafts/
  faq-approved/
  faq-promotion-log.md
  faq-parity-scan-2026.md

Link folder from release-evidence taxonomy.

How this differs from adjacent workflows

Run disclosure challenge after FAQ pipeline produces approved files—or merge approved FAQ into Day 2 store sweep.

Sample approved FAQ rows (pattern only)

These illustrate bullet → approved answer shape—not copy for your game.

Source bullet: AI: NPC barks use cloud LLM; listed in ai-disclosure annex
Approved Q: Does the game use AI-generated content?
Approved A: Yes. NPC dialogue barks are generated at runtime using a cloud language model. Details appear in our AI disclosure annex and store About section. Voice acting for main story scenes is human-recorded.

Source bullet: NOT in demo: crafting
Approved Q: Is crafting in the Next Fest demo?
Approved A: No. Crafting appears in the full game roadmap; the festival demo stops after chapter 1 and does not include crafting systems.

If the model drafts "crafting-lite in demo," Gate D deletes it—no new bullet without producer scope change.

Question bank starter (micro-studio)

Generate drafts per ID—not one blob—so promotion logs stay granular.

CI and pre-commit hooks (optional)

Mirror spirit of validate-packet exit codes—FAQ promotion can use exit 1 script if faq-promotion-log.md lacks today's build id.

Partner diligence export bundle

When assembling diligence packets, zip:

1. faq-source-bullets-v*.md
2. faq-approved/ markdown or PDF export
3. faq-promotion-log.md
4. faq-parity-scan-2026.md
5. Link to ai-disclosure/ annex
6. Screenshot of live Steam FAQ section with build id in filename

Filename: faq-governance-2026-05-18-build-rc7.zip aligned with partner ZIP naming.

Asia-EU handoff note

Distributed teams often draft FAQ in one timezone and publish in another. Add to handoff evidence checklist:

• EU morning: parity scan on overnight build
• Asia evening: no FAQ publish without log row signed in shared doc

Prevents "Friday night LLM paste" without Monday review.

Hallucination catalog (delete on sight)

Train reviewers on catalog once—reuse every Gate D.

Weekly rhythm during Q3 prep

Avoid FAQ rewrites on same day as major demo depot promotion—two change vectors confuse parity scan.

Red-team pass (optional, 60 minutes)

Before partner zip:

1. Reader with no project context reads FAQ only.
2. List claims they believe are promises.
3. Compare to demo install on fest demo depot.
4. File mismatches as RED in parity scan.

Matches mock audit tabletop tone.

Steamworks paste discipline

Steam FAQ fields are not git—treat paste as promotion event:

1. Copy from faq-approved/Q03_ai.md only—not from chat UI.
2. Paste into Steamworks; screenshot result.
3. Add screenshot path to faq-promotion-log.md.
4. Cold-read FAQ on live store page in incognito.

If Steam formatting strips bullets, adjust approved markdown once—do not improvise in Steamworks text box.

Mobile and multi-storefront parity

Games with two-storefront rule still need one source bullet file. Secondary store may shorten answers—run Gate E per surface; log separate promotion rows (Q03_ai_steam, Q03_ai_secondary).

Mobile forms may use different AI questions—map answers to same bullets, not separate model sessions.

Intake compression and 72-hour recovery

When intake compression triggers yellow on FAQ mismatch:

1. Pull last approved FAQ from git tag.
2. Diff against live store.
3. Fix copy or fix build—do not argue with partner while FAQ still lies.
4. Attach recovery diff to hash case study-style evidence folder (pattern, not same failure class).

Operating review integration

Teams using four-Friday operating reviews add FAQ gate status to Block 2 (store/marketing):

RED blocks fest marketing cap increases.

Tabletop rehearsal (30 minutes)

1. Producer reads FAQ aloud.
2. Engineer says pass/fail per bullet live on demo.
3. Marketer checks 32px icon and FAQ do not contradict genre promise.
4. Sign parity scan or file RED list.

Localization and translated FAQs

Machine translation of approved English FAQ is another Gate D:

1. Export faq-approved/ after English Gate E passes.
2. Translate with human post-edit or locked MT workflow.
3. Native reader signs faq-promotion-log.md row per locale.
4. Never run Gate C directly into non-English without English source freeze.

See localization QA tools for string checks—FAQ is storefront copy, not just in-game strings.

Fail pattern: English FAQ honest; German FAQ implies features English denied—regional partner tests both.

Versioning when bullets change mid-cycle

Tag git commits: faq: promote Q03_ai for build rc7—searchable during audits.

Metrics you can track (without inventing industry benchmarks)

Improve process; do not optimize for "fastest LLM."

Beginner quick start (45 minutes)

1. Write ten source bullets in plain English.
2. Paste into locked template; generate one answer.
3. Red-pen anything not in bullets.
4. Ask teammate to sign one line in promotion log.
5. Stop—do not publish until Gate E exists.

Success: you can point to a bullet for every FAQ sentence.

Common mistakes (seven)

1. Skipping Gate A — Model becomes product owner.
2. Same person generates and signs — No effective review on AI rows.
3. Publishing draft file path — Wrong markdown ships.
4. FAQ promises fest features not in demo — Truth audit fail.
5. Template changes without re-diff — Hidden tone/claim drift.
6. Ignoring regional price FAQ — Conflicts with pricing worksheet.
7. FAQ AI story ≠ runtime registry — Partner catches prompt hash mismatch.

Pro tips (six)

1. Question IDs stable — Q03_ai across versions for clean logs.
2. Store FAQ in git — Real diffs, not Google Doc mystery edits.
3. Link FAQ promotion to validate-packet exit 0 before partner upload week.
4. Refresh FAQ when demo depot changes — Depot audit date triggers FAQ Gate E.
5. Cap LLM refreshes — One batch per evidence cycle, not daily rewrites.
6. Export PDF of approved FAQ — Some partners want non-git artifact.

Snippet-friendly answers

Can we use ChatGPT to write our Steam FAQ?
Yes for drafts inside this pipeline; no for direct publish without Gate D human diff and Gate E parity scan.

How is this different from legal review?
This is factual parity and AI honesty discipline—not substitute for counsel on refunds, privacy, or regional law.

Do we need this if we have no LLM in the game?
If FAQ is 100% human-written, skip Gates B–C; still run A, D (self-diff), and E.

Key takeaways

• Store FAQs are binding copy for partners and players in 2026 reviews.
• Five gates: frozen bullets, locked template, LLM draft, human diff, parity scan.
• Models rephrase; humans own facts via source bullets.
• High-risk topics need two human sign-offs.
• Log every promotion in faq-promotion-log.md.
• Align with prompt registry and disclosure challenge.
• Export under release-evidence/ai-disclosure/faq-pipeline/.
• Fest teams tie FAQ refresh to demo depot and marketing cap discipline.
• Speed without diff gates is liability, not leverage.
• Translated FAQs need their own Gate D pass—never auto-translate unreviewed English drafts.
• Git-tagged FAQ promotions survive 72-hour intake recovery windows.

FAQ

What tools for human diff?
Git diff, Meld, VS Code compare—pick one; consistency matters more than brand.

Can one founder team satisfy two-human rule?
Use external advisor or publisher contact for high-risk sign-off—document role in log; do not fake a second internal hat.

How often refresh FAQ before October fest?
When build_id changes player-visible scope; otherwise monthly scan during UTM experiment period.

Does FAQ pipeline replace AI disclosure challenge?
No—import approved FAQ answers into Day 2–3 sweeps.

What if model outputs perfect prose?
Still diff—perfection is how hallucinations hide.

Can we run FAQ pipeline in parallel with cold-hash challenge?
Yes—different owners; same release week should show both green in evidence folder.

Does FAQ need hashing?
Optional: sha256 of faq-approved/ folder in BUILD_RECEIPT annex for partners who ask for integrity proof—use same tools as validator listicle. Hashing is optional; human diff is not.

Conclusion

Q3 2026 partner review will not grade your prompt engineering—it will grade whether store words match the build. The FAQ LLM draft pipeline keeps speed and adds the diff discipline micro-studios skip when tired.

Freeze bullets this week. Lock the template. Let the model draft. Make a human diff the promotion gate—not the publish button.

Save artifacts beside your registry freeze and disclosure annex. When the partner asks "who approved FAQ row 7," you open a log—not a chat history.

The pipeline's point is simple: LLMs draft; humans promote. Anything else treats the model as co-founder—and partners noticed that failure mode often in 2026 intake cycles. Start with five source bullets and one FAQ answer this week; scale the folder when it hurts less than rewriting store copy in a panic. Export the promotion log the same day you paste to Steam—future you will not remember who signed off during October fest week chaos.