# RobCo U.O.S. — Prompt Library > A reusable, discipline-tagged toolkit of "engineering-campaign" prompts for running Dispatch (Fable 5 / Opus 4.8 / Sonnet 5) against this repo. > **Living document** — grow it as new campaigns prove useful. Gitignored (local only, never on GitHub), like all of `planning/`. > Built 2026-07-04 from the parked audit chain + a review of external prompt suggestions. See memory `prompt-library-campaign-framings`. --- ## How to use this 1. **Pick a campaign, not a task.** These are 1–12 hour repository-wide investigations, not "add feature X." The limiting factor isn't coding ability (Dispatch has all three models + repo context + 2000+ tests) — it's finding work worth doing. 2. **Analysis before code, always.** Every campaign produces a triaged report FIRST (finding → why it matters → impact → dependencies → risk → verdict: NOW / DEFER / REJECT). Only after the report, implement the highest-value items in verified phases. 3. **Feed synthesis.** Reports land in `planning/` and feed the master-plan synthesis so campaigns compound instead of colliding. 4. **House rules are law in every campaign** (they're already in RULES.md/CLAUDE.md; restate the load-bearing ones in the prompt): the full 2000+-test gate is law (never `--no-verify`); both test runners at exact parity (Protocol 15); Protocol 1 cache bump on served-file change; Protocol 2/2a docs+count sync; dev-branch only (Protocol 43); UTF-8 safe writes (Protocol 39); game-agnostic (Protocol 38); player-control principle; **modernize WITH immersion, never strip modern features (design principle 0j)**. 5. **Don't over-audit.** The failure mode is analysis-paralysis — more reports than shipped code. Run the few highest-value campaigns, implement, then re-baseline. (Protocol 28 usage efficiency.) ## Reusable prompt scaffold (paste + fill the `<CAMPAIGN>` line) ``` Run a repository-wide <CAMPAIGN> on RobCo U.O.S. (analysis first, no code until the report is done). Read the whole repo — architecture, js/, css/, tests, docs, the CLAUDE/RULES protocols, and planning/ history — and build a mental model before proposing anything. <CAMPAIGN-SPECIFIC FOCUS>. Challenge existing assumptions rather than preserving them by default. Produce a prioritized report to planning/<NAME>_AUDIT.md: each finding → why it matters → expected impact → dependencies → risk → verdict (implement NOW / DEFER / REJECT), severity-ranked, per-file with a file-touch list. Constraints: behavior-neutral where the campaign is a refactor; full gate is law + both runners at parity; game-agnostic; UTF-8-safe; dev only; never bump APP_VERSION without owner sign-off. Implement NOTHING yet — report, then stop. ``` --- ## Workflow techniques (adopted 2026-07-04) ### Red-Team / Devil's-Advocate pass (for big design/architecture calls) Extends the Protocol-8 plan→audit with an adversarial stage. Use for high-stakes decisions, not routine work. ``` Take the attached plan/design for <X>. Before we build it, RED-TEAM it: argue the strongest possible case AGAINST it — where it fails, what it breaks, what a senior engineer / a Fallout purist / a mobile user would hate, cheaper alternatives, and what future problem it bakes in. Then reconcile: what survives, what changes, what's rejected. ``` (In practice: Opus argues against → the plan is revised → Sonnet builds whatever survives.) ### Open-ended "no objective" campaign (run ONCE per round, as fresh eyes) Time-boxed, analysis-only, feeds synthesis. Catches blind spots the structured audits miss. ``` Perform a repository-wide engineering campaign with NO predefined objective other than making RobCo U.O.S. objectively better. Read everything — architecture, implementation, tests, docs, protocols, workflows, planning docs, history. Build a mental model before proposing changes. Find: architectural weaknesses, inconsistencies, simplifications, immersion improvements, technical debt, missing abstractions, underused systems, scalability risks, Fallout-authenticity issues, UX friction, performance bottlenecks, and emergent-feature opportunities where existing systems could combine. Challenge assumptions. Rewrite the roadmap if implementation order should change. Produce a prioritized report (why / impact / dependencies / risk / now-defer-reject). Report only — do not implement. ``` Variants (same scaffold, different lens): **Surprise Me** (find improvements I'd never think to ask for) · **Embarrass Me** (everything an experienced engineer would criticize) · **Forget the Roadmap** (ignore planned work — what actually deserves attention?). ### Session-end capture (close of any substantive session) — adopted 2026-07-23 The honest fix for **conversation→queue drift** (G review, 2026-07-22): a script can't read a conversation, so a decision made and acted on in-session slips the automated memory↔queue nudge (Protocol 50). Close it with a ritual, not a guard. *(Curated in from Anthropic's Claude Code prompt library — the one tactical prompt there that rises to a reusable technique for us.)* ``` Summarize what we actually did and decided this session — shipped changes, decisions reached, reversals, anything deferred or newly discovered. Then tell me exactly what to write WHERE: which QUEUE.md items to update/close, which orchestrator memory to add or amend (decisions + reasoning, NOT lookup-able facts), and whether CLAUDE.md/rules need a line. Propose the edits; don't guess — flag anything you're unsure belongs. ``` RobCo's version writes to THREE stores (queue + memory + rules/CLAUDE.md), not just CLAUDE.md — and memory holds reasoning/decisions, never facts you can look up. Pairs with Protocol 50. --- ## Disciplines Each maps to an existing parked audit where noted (don't duplicate — fold the framing into that audit). ### 01 · Architecture → parked `robco-audit-program` / synthesis - **Repository archaeology:** read every file; determine WHY every system exists; find systems that solved old problems but are now obsolete; recommend deletions / simplifications / consolidations. Report before touching code. - **Missing abstractions:** find repeated shapes that want a shared helper/manager (respecting Protocol 22 extend-before-create + Protocol 23 layering). - **Load-order & coupling map:** diagram cross-file coupling vs the documented script load order; flag violations. ### 02 · Technical Debt → parked `parked-code-quality-audit` - **Delete code:** reduce repo size without reducing capability. - **Magic numbers:** find every unexplained constant; centralize + document + name. - **Naming audit:** variables / functions / files / folders — consistency pass. - **Readability audit:** reduce cognitive load / complexity (not formatting). - **CSS split question:** should `css/terminal.css` be split like `ui.js` was? (already logged in the code-quality audit.) ### 03 · UX → parked `parked-mobile-audit` + UI-consistency - **Friction audit ("think like Apple"):** reduce clicks / taps / friction per workflow. - **Frustration audit ("think like Nintendo"):** find and remove annoyances. - **Empty/error/loading-state audit:** every state has intentional, in-voice copy. ### 04 · Fallout Immersion → parked `parked-diegetic-audit` (derives HOUSE_STANDARD) - **Terminal personality:** does every message sound like RobCo? Can it? - **AI personality:** does every AI response belong inside Fallout? - **CRT/convention audit:** which modern conventions remain — and which SHOULD stay (per 0j) vs which have a better terminal-native form? (NOTE: modernize WITH immersion; do NOT strip modern features.) - **Environmental storytelling:** can the terminal tell stories by just existing (boot logs, old records, hardware history) — no quests/dialogue? - **Boot-sequence expansion:** diagnostics, warnings, old logs, hardware checks, radio chatter. - **Terminal history:** what happened before the player owned this machine? - **Corporate bureaucracy:** replace flat UI moments with in-world RobCo bureaucracy where it improves immersion without hurting usability. ### 05 · AI → parked `parked-ai-feature-eval` + `parked-ai-to-native-feature-audit` - **AI→native audit:** every AI-dependent feature that could be deterministic/offline → make it native (player-control principle). - **AI-first architecture:** if every future feature is AI-assisted, how should the architecture evolve? ### 06 · Performance → parked `parked-performance-audit` - **"Think like Carmack/id":** optimize render/memory/hot-paths; no feature additions. ### 07 · Security → existing security suites + firestore rules - **Threat-model pass:** injection surfaces, cloud rules, kill-switch fail-safety, App Check. ### 08 · Testing → parked `parked-test-strength-audit` - **Test-strength audit:** are the 2000+ tests meaningful or just present? Find weak assertions, coverage gaps, runner-parity risks. ### 09 · Accessibility → parked `parked-accessibility-audit` - **A11y campaign:** contrast, reduced-motion, focus, tap targets, screen-reader semantics — beyond the automated baseline. ### 10 · Code Quality → parked `parked-code-quality-audit` - (see 02.) ### 11 · PWA - **Install/offline/update audit:** service-worker invariants, precache completeness, update-prompt path, standalone-mode edge cases. ### 12 · Audio → parked capability survey - **Sound-design audit:** is every interaction's audio intentional + mutable + dial-gated (Protocol 7)? Gaps + opportunities. ### 13 · Worldbuilding → parked `parked-diegetic-audit` + gameplay ideation - (see 04 environmental storytelling / terminal history.) ### 14 · Data Integrity → parked `parked-cloud-audit` - **Round-trip audit:** every state field survives save → sanitize → migrate → apply → cloud; no silent drops. ### 15 · Automation - **CI/gate self-improvement:** every escaped defect ratchets the gate finer (Protocol 36b) — find classes not yet guarded. ### 16 · Documentation - **Doc-accuracy audit:** README/ARCHITECTURE/CLAUDE vs actual code; stale counts, dead references. ### 17 · Future-proofing → synthesis - **Predict future problems:** imagine RobCo at v5.0 — which decisions today become debt? - **Museum / Modding / Community / AI-first audits:** how should architecture change if preserved 20 years / extended by strangers / hit by 500 contributors / all-AI features? ### 18 · Parallelization → parked `robco-audit-program` (parallelization plan) - **Worktree parallel-execution plan:** run ON the master-plan backlog → PARALLELIZATION_PLAN.md. ### 19 · Game Feel → gameplay round (2.9) - **Cross-System Interaction audit ⭐:** find systems that never talk but SHOULD (timeline/GPS/inventory/quests/companions/faction/vendors/weather/time/audio/AI). The best emergent features live here. - **Polish ("think like Valve"):** micro-interactions, responsiveness, feedback. ### 20 · Polish - **Final-mile audit:** the small stuff — alignment, spacing, motion consistency, copy tone — against the house standard. --- ## Notes - **Scrapped:** "What would Bethesda do?" (strip everything that feels like a modern website) — conflicts with design principle 0j. The good kernel (audit interactions for immersion) is 04's terminal/AI-personality + CRT audits, framed as modernize-WITH-immersion. - **Where these run:** the planning engine runs at the FRONT of each round (audits → synthesis → master plan → parallelization). Enrich the round's audit prompts with the framings above; pull individual prompts from here any time. - **Keep it lean:** run the few highest-value campaigns, implement, re-baseline. Don't generate a giant library for its own sake.