# Research Development Life Cycle v1.4.3 | updated 2026-07-31 | [CHANGELOG](./CHANGELOG.md) A product design and research methodology, packaged as an agent skill. Research Development Life Cycle (RDLC) governs everything from the first signal to a validated design: sensing problems in your research corpus, discovery, design definition, ideation, prototyping, and human validation. Its terminal artifact, the **Validated Intent**, is the input handed to whichever engineering process builds it next. > AI proposes, humans decide. Evidence before opinion. Failure paths before > happy paths. Questions adapt to the host: capable terminal interfaces use arrow-key selection and Enter, desktop interfaces use option buttons, and every question still accepts a selected option plus context or a fully open-ended answer. ## Phase map ``` SIGNALS ─► DISCOVERY ─► DEFINITION ─► DESIGN ─► VALIDATION ─► [engineering] ─► OPERATIONS (mine the (is it real, (mental CONSTRUCTION (trust Inception → (post-ship corpus) is AI right) model, (prototype scorecard, Construction listening, failure loops, HAX ship gate) → Operations feeds SIGNALS) modes) checks) ``` ## Start in any AI app Open ChatGPT, Claude, Gemini, or another AI app with web access and paste: ```text Fetch and follow RDLC Conversation Mode at https://rdlc.org/rdlc-rules/chat-workflow.md. Help our team decide what product work is worth doing. Work in this chat. Do not ask us to install anything or use a command line. Ask one decision-shaping question at a time and keep a compact RDLC Working Record. Our situation: [Describe the product idea, research, prototype, or decision.] ``` Conversation Mode keeps evidence, assumptions, decisions, risks, conflicts, and the next action in one copyable Working Record. If the AI app cannot fetch a URL, use the "Copy with the method" button at https://rdlc.org. ## Workspace setup Use this path when a coding agent should write persistent, versioned RDLC artifacts into a repository. ```bash git clone aidlc-design-workflows cd your-project # install steering rules for your agent (claude | cursor | copilot | cline | kiro | q | agents) ../aidlc-design-workflows/scripts/setup.sh claude # feed the corpus (de-identified!) then inventory it cp ~/exports/*.csv research-corpus/support-tickets/ python3 scripts/corpus/ingest.py ``` Then start your agent and say: ``` Using RDLC signals, mine the research corpus and surface our top unframed problems. ``` or, with a problem in hand: ``` Using RDLC, help us decide whether to build . ``` Trigger phrases: `Using RDLC, ...` (adaptive) plus phase jumps: `signals` / `discovery` / `define` / `prototype` / `validate` / `listen` / `kb`. ## What it enforces - **Risk tiers (T1 regulatory → T4 internal)** scale every gate. Payments and personal-data work cannot skip validation; internal tools ship fast. - **Mental model before pixels.** No prototyping until the user's mental model of the AI is written and approved. - **Failure before happy path.** Wrong-AI states are designed first; every validation study includes a mandatory failure-recovery task. - **Trust Scorecard ship gate.** Comprehension, Control, Failure Recovery, Trust Delta, with thresholds by tier. Engineering KPIs measure whether the software works; this measures whether humans trust it. - **AI necessity test.** Sometimes the right answer is a better form, not a model. The workflow says so. - **Two working modes.** Conversation Mode works inside an ordinary AI chat with no install. Workspace Mode writes persistent project artifacts for resume, handoff, and version control. - **A living knowledge base.** In Workspace Mode, `rdlc-docs/kb/` holds business context, market context, personas, a value map, numbered learnings, decisions, and conflicts. Every phase gate ends with a mandatory KB sync; freshness sweeps flag stale personas; nothing is approved with un-synced learnings. - **Strategist lens.** Elicitation banks interrogate the business, the users, and the market; every intent must state user value, business value, and the causal linkage between them. A "hopeful" linkage blocks T1/T2 ship gates. - **Conflict and challenge protocol.** Contradictions between artifacts, KB entries, and evidence are detected at gates and logged as C-NNN with forced resolution. Unexamined assumptions get exactly one plain challenge; the user's decision is then logged with the overruled evidence. - **Design Units of Work.** Work splits by the question it answers, one owner each, with interface contracts, crit rituals that cite principles by number, and calibration from solo designer to multi-squad teams. - **Session continuity, handoff, and splitting.** A live `state/` folder (resume.md, an append-only session log, per-unit worklogs with OPEN THREADS) lets any designer or a fresh agent resume cold, take over a unit, or split work in parallel, with no synchronous handoff meeting required. `Using RDLC resume` / `handoff DUW-NN to ` / `split `. - **Prototype construction, grounded in real solutions.** CONSTRUCTION starts by asking where the prototype begins (green field, existing Figma, live page, repo, or a Chrome-extension emulation of a live product) and whether a design system exists. It reviews every loop live via Chrome MCP, citing named UX laws and design principles rather than unattributed taste, debates genuine tensions explicitly instead of resolving them silently, and gives an explicit READY / NOT READY verdict before entering VALIDATION. Loop-backs from VALIDATION re-enter as scoped iterations grounded in what specifically failed. See `common/ux-laws.md` and `docs/CHROME-MCP-SETUP.md`. - **External research on demand.** When an idea or thin-market problem is on the table, RDLC offers competitive analysis, market and category scans, recent news, analogous patterns, and regulation/capability context, and runs it (if a search tool is available) only when you say yes. Findings enter the KB as sourced, numbered learnings, never as unattributed confidence. - **The 5 Universal AI Agents.** Research Synthesizer, Competitor Intelligence, Problem Definition, Idea Generation, and UX Flow Designer (Nick Babich, "Agentic Product Design," UX Planet) map directly onto RDLC's SIGNALS/DISCOVERY, external-research, DISCOVERY, IDEATION, and DEFINITION stages, contributing named failure-mode checks, a friction-type taxonomy, and explicit human-in-the-loop boundaries at each. See `common/design-agents.md`. - **A killed/ folder.** Stopping on evidence is logged as a success. ## NotebookLM / corpus mining The SIGNALS phase mines a local `research-corpus/` (interviews, tickets, NPS verbatims, sales notes, analytics) and can also mine NotebookLM / Gemini Notebook directly. Three paths, detailed in docs/NOTEBOOKLM-INTEGRATION.md: (A) live MCP via community servers such as notebooklm-mcp-cli, letting the agent query notebooks with citations; (B) CLI batch pulls using the shipped prompt files; (C) manual export of the four mining prompts. There is no official consumer API yet (Enterprise API exists via Google Cloud), and Paths A/B use unofficial cookie-auth tooling, so the doc includes a data-governance gate: NDA'd or personal material goes through security review or stays on Path C. All mined answers land in `research-corpus/notebooklm-exports/` so evidence stays file-based, git-audited, and PII-free. ## Repo layout ``` rdlc-rules/ core-workflow.md # the steering file agents load rdlc-rule-details/ # lazily loaded per-phase rules common/ # risk tiers, principles, KB, conflict protocol, # strategist lens, work units, session continuity signals/ discovery/ definition/ design-construction/ validation/ operations/ signals/prompts/ # NotebookLM/CLI mining prompt files extensions/ # payments-risk, personal-data-trust (opt-in, blocking), # accessibility (always on) scripts/ setup.sh # installs rules for claude/cursor/copilot/cline/kiro/q corpus/ingest.py # normalizes + inventories the corpus research-corpus/ # your evidence goes here (de-identified) docs/ # artifact reference, working guide, NotebookLM setup # generated at runtime in your project: rdlc-docs/ kb/ # living knowledge base state/ # resume.md, session-log.md, worklog/ <- resume & handoff signals/ discovery/ definition/ ideation/ design-construction/ validation/ operations/ plan/ questions/ killed/ for-engineering/ ``` ## Extensions Same mechanism as engineering: `*.opt-in.md` files present a question during Problem Framing; enabled extensions become blocking rules with verification criteria (PAY-01..04, PDT-01..04, A11Y-01..03). Add your own by dropping a rules file + opt-in file into `extensions//`. ## Relationship to engineering | | engineering | RDLC | |---|---|---| | Question | Are we building it right? | Are we building the right thing? | | Cadence | Bolts | Prototype loops | | Collaboration ritual | Mob Elaboration / Construction | Framing sessions, study reviews | | KPIs | Deployment, recovery, defect metrics | Comprehension, Control, Trust Delta | | Output | Working software | Validated Intent (engineering's input) | During later engineering stages, run the reduced loop (HAX check + content review) inside each Bolt for any Unit of Work with a user-facing surface. ## License MIT-0, matching the upstream engineering workflows repo convention.