ai-native-conversational-os-design
Designs the conversational UI architecture for an AI-native product in six stages — Intent, Strategy Cards, Clarification, Output Cards, Mission, Agent Execution — applying five AI-first product princ
SKILL.mdspecialisation-packs/ai-native-startup-design/skills/ai-native-conversational-os-design/SKILL.md
GitHubAI-Native Conversational OS Design
Purpose
Provide a concrete, reusable architecture model for how an AI-native product is built as a UI that is NOT a traditional collection of screens/menus/dashboards but a conversational operating system: the user's intent is recognized, the right internal "strategy" is selected, clarifying questions are asked when needed, structured output cards are produced, one clear mission is given, and an agent can continue the work autonomously. Core message: "your product is no longer a set of screens. It's a thinking partner."
Based on
- The methodology of an external "AI-first SaaS Product" workshop,
applied by the owner to one own case
("Decision Coach" MVP) — see
../../references/ai-first-saas-workshop-source.mdand the worked example../../cases/ai-decision-coach-mvp-case.md, section 8. Note: applied only once so far — not broadly validated across multiple different products. - The workshop's "5 shifts" principles for designing an AI-first product (see the start of the Method section).
Method (draft — to be filled in further)
A. Five AI-first product principles (mindset before architecture)
Before designing the OS flow, internalize these five shifts from old SaaS thinking to AI-native thinking:
- Click → question. The user doesn't navigate menus to find the right function — they ask for what they want, and the system finds the right function.
- Menus → prompts. Instead of fixed menu structures, the user expresses their intent in natural language.
- Dashboards → dialogue. Instead of browsing information, information is brought to the user through conversation, at the right time, in the right context.
- Manual actions → agents. The user doesn't perform every step
themselves — an agent performs them, the user directs and approves
(see
../closed-loop-process-and-human-oversight-design/SKILL.mdfor choosing the level of human oversight). - Screens → chat + cards. The UI isn't a fixed collection of screens but a dynamic combination of conversation and structured information cards that appear as needed.
The shared conclusion of these five shifts: the product is no longer a set of screens, it's a thinking partner.
B. Six-stage OS flow
- Intent (user → system). Identify WHY the user is here and what they want clarity on. Explicitly list the main intents your product supports (typically 3-6) — don't try to support an unlimited range of free-form requests in an MVP. Identify the dominant intent and pass it to the strategy layer.
- Strategy Cards (system → internal reasoning layer). Define
"playbooks" (strategy cards) the AI can choose from based on the
user's intent. Each card is an independent reasoning module: what it
interprets, what it produces (e.g. a 0-100 score, a classification, a
reworded text). Design as many cards as the MVP's differentiator and
table-stake needs require (see
../customer-vision-to-jtbd/SKILL.md) — no more. - Clarification (interactive micro-questions). Ask AT MOST 2-4 clarifying questions, only when (a) the input is too vague to interpret, or (b) the wrong strategy card has been activated. Keep the questions light and fast — this isn't a form, it's a refinement.
- Output Cards (core MVP results). Design standardized, structured card formats in which the user receives the result of each strategy card execution (e.g. a score + a "why this score" rationale + "what would improve it"). Each output card should directly fulfill one of the MVP's differentiator or table-stake needs.
- Mission (AI summarizes the plan + the next step). One short mission statement at the end of the session that frames the next steps around building trust and reducing uncertainty — not a long summary, but one concrete, action-driving sentence.
- Agent Execution (system → autonomous action). After the mission statement, an agent can continue independently: updating scores as new information arrives, rewriting material, recommending existing tools/resources. The agent's job is to create forward momentum — not just answer a question and stop.
C. Design checklist
- Test the flow end to end before building. Write out one concrete user journey from the Intent stage to the Agent Execution stage in words (not code) — if any step feels forced or artificial, simplify the structure before the build phase.
- Feed the flow into
../ai-buildable-prd-writing/SKILL.md's "Core Features" section — each stage of the OS flow (Strategy Card, Output Card) is one PRD feature line, described as an outcome ("the user gets...") rather than a technical implementation.
What this skill does NOT do
- Does not include a technical orchestration implementation (prompt
chaining, state, API interfaces) — it produces the conceptual
architecture, which is handed to the build agent via the tool chosen
through
../ai-native-tool-stack-selection/SKILL.md. - Does not replace
../closed-loop-process-and-human-oversight-design/SKILL.mdfor deciding the human oversight level for the Agent Execution stage — use it alongside this skill to decide the in/on/outside-the-loop level for each agent action. - Does not fit every product — if the product is genuinely tool-/dashboard-type (e.g. data visualization, continuous monitoring without conversational decision-making), this model forces the wrong shape. Use it only when the core value is AI interpretation/reasoning, not displaying data.
Refinement notes
This skill has so far been applied to one case (the owner's Decision Coach). As you apply it to more products, add:
- your own observations on when the 6-stage model needs to be simplified (e.g. if the Strategy Cards layer proves oversized for a small MVP)
- concrete examples of other OS flow designs in the
../../cases/folder - observations on how the flow performed in practice after the first build iteration (which stage produced the most user value, which proved unnecessary)
Once this section has been filled in with multiple cases, raise
skills_index.json's maturity field to validated
(see ../../../../meta/maturity_levels.md).
Continue from here
- Preceding skill in this pack:
../rice-scoring-and-mvp-synthesis/SKILL.md— produces the chosen MVP for which the OS flow is designed. - Next skill in this pack:
../ai-buildable-prd-writing/SKILL.md— feeds the OS flow into the PRD's Core Features section. - Related skill in this pack:
../closed-loop-process-and-human-oversight-design/SKILL.md— choosing the human oversight level for the Agent Execution stage. - Related skill in another pack:
../../../../business-design-frameworks/skills/customer-journey-and-ai-touchpoint-mapping/SKILL.md— a complementary way to structure the same product as a customer journey rather than an OS architecture. - Worked example:
../../cases/ai-decision-coach-mvp-case.md, section 8. - The pack's shared guardrails:
../../CLAUDE.md
References
../../references/ai-first-saas-workshop-source.md— source information../../cases/ai-decision-coach-mvp-case.md— worked example../../CLAUDE.md— the pack's shared guardrails