claude-rdlc-wizard
June 10, 2026 · View on GitHub
RDLC enforcement for Claude Code — hooks, skills, and wizard setup for research repos. The research-domain sibling of claude-sdlc-wizard.
Status: v0.1.0 (bootstrap, 2026-05-04). Three case studies cleared the xdlc "two case studies → extract" threshold: anticheat, states-project-research, tucson-investigation. Pattern catalog lives in PATTERNS.md; proof-points in CASE_STUDIES.md; build journal in EXTRACTION_NOTES.md. The standalone ~/rdlc/ repo was consolidated into this wizard on 2026-05-04 (mirrors GDLC retirement pattern).
What this installs
A research lifecycle on top of any Claude Code project that produces sourced research deliverables — interview prep, medical/legal evidence dossiers, automotive diagnostics, market analyses, anything where claims need sources and confidence levels.
| Layer | What it does |
|---|---|
RDLC.md | Consumer-installable canonical doc — confidence vocabulary, source hierarchy, slop gate, audience-firewall rule |
/rdlc skill | Full lifecycle workflow (Plan → Verify → Review → Ship → Improve) |
/setup-rdlc skill | Confidence-driven setup — scans repo, asks only what it can't detect |
/update-rdlc skill | Drift reconciliation when wizard upgrades |
/feedback skill | Privacy-first contribution loop |
| Hooks | Slop scan, confidence-required, source-required, prompt-check, audience-firewall |
| Templates | regression_test.sh, slop_scan.sh, multi-deliverable generator scaffold |
Install
curl -fsSL https://raw.githubusercontent.com/BaseInfinity/claude-rdlc-wizard/main/install.sh | bash
Or scaffold manually:
npx -y claude-rdlc-wizard init
Run inside a research repo. The wizard scans for evidence/sources/confidence-marker patterns and only asks what it can't infer.
Relationship to claude-sdlc-wizard
Both run in parallel. SDLC handles code quality; RDLC handles research correctness. From ~/xdlc/README.md:
SDLC without the domain DLC = clean code that produces wrong output. Domain DLC without SDLC = correct requirements built with sloppy code.
A research repo installs both:
curl -fsSL .../claude-sdlc-wizard/install.sh | bash # code layer
curl -fsSL .../claude-rdlc-wizard/install.sh | bash # research layer
Vocabulary mapping (vs SDLC)
| SDLC | RDLC |
|---|---|
| Write failing test | Form hypothesis / define claim |
| Implement feature | Gather evidence / research |
| Run tests (green) | Cross-validate sources |
| Lint / typecheck | Confidence scoring |
| Regression tests | Fact regression tests |
| Code review | Cross-model adversarial review |
| TDD guard hooks | Source verification hooks |
What's in v0.1.0 vs deferred
In v0.1.0:
- Skill triple (
rdlc,setup,update,feedback) - Five hooks (prompt-check, slop-scan, confidence-required, source-required, audience-firewall)
- Three templates (regression_test.sh, slop_scan.sh, generate_deliverable.py skeleton)
- RDLC.md consumer canonical
- install.sh + package.json
Deferred to later versions (per skills-first rule, earned through use):
- Setup wizard scan logic refinement (starts with the 5-row signal table from xdlc cross-domain-concerns.md)
- Per-domain RDLC presets (medical/legal, political, automotive, journalism)
- Codex adapter (
codex-rdlc-wizard) - L-code enumeration (anticheat A–G, tucson L1–L15 — earn RDLC's own through use)
- npm registry publish (currently local install via
npx --prefixagainst the cloned repo)
Documentation
RDLC.md— the consumer canonical that gets installed into target reposPATTERNS.md— pattern catalog and lessons learned across case studiesCASE_STUDIES.md— proof-point cross-index + Contributions Inventory (artifact → wizard slot)EXTRACTION_NOTES.md— build journal from the v0.1.0 extractionWIZARD_PLAN.md— historical implementation plan (superseded by v0.1.0 ship)HANDOFF.md— historical pre-build handoff (superseded)CLAUDE.md— wizard self-instructions (this repo)ARCHITECTURE.md— how the pieces fitCHANGELOG.md— version historyROADMAP.md— what's queued for v0.2 and beyond
Case studies that proved the patterns
See CASE_STUDIES.md for the full cross-index. Short version:
- anticheat (160 commits, 316 tests) — medical/legal evidence aggregation. GRADE-aligned labels, certification queue, compound mechanism verification.
- states-project-research (43 commits, 257 bash asserts) — political/interview prep. VERIFIED/SUPPORTED/INFERRED/UNVERIFIED, fact regression suite, multi-deliverable generator with audience firewall.
- tucson-investigation (36 commits, 111 tests) — automotive diagnostic + dealer audit. DIRECT/SUPPORTED/INFERRED/GAP, Playwright persona tests, methodology leak gate, 9-round Codex xhigh loop.
License
MIT
AI Setup Lanes
Three recommended setups in AI_SETUP_LANES.md:
| Lane | Planner | Driver | Reviewer | When |
|---|---|---|---|---|
| A — Research Premium | Opus 4.6 max | Opus 4.6 max | GPT-5.5 xhigh | Research drafting, confidence-critical, source evaluation |
| B — Research Saver | Opus 4.6 max | Sonnet (latest) | GPT-5.5 xhigh | Routine drafting, formatting, docs |
| C — Research Lite | You | Haiku 4.5 | None | Data formatting, bibliography, file management, grunt work |