Skip to content

issue-013.md

Latest commit

 

History

History
89 lines (54 loc) · 5.95 KB

issue-013.md

File metadata and controls

89 lines (54 loc) · 5.95 KB
issue_type Feature
linear_workflow New product surface — chat-first research copilot (not a fixed command pipeline); implementation app/repo path TBD after `/explore` + `/create-plan`
source PM brief (2026-04-12) — conversational research agent with planning, orchestration, and export
prior_cycles N/A (greenfield product direction in AI Product OS)
linear_root VIJ-65 — https://linear.app/vijaypmworkspace/issue/VIJ-65/issue-013-pm-research-copilot-chat-first-planning-orchestrated
linear_project https://linear.app/vijaypmworkspace/project/issue-013-pm-research-copilot-chat-first-planning-orchestrated-a5e43bf29c24
app TBD (portfolio-level; target codebase after exploration)

Issue Title

PM Research Copilot — chat-first planning, orchestrated evidence, exportable insights

Type: Feature

Problem (Current State → Desired Outcome)

Current state: Many PM workflows are still pipeline-shaped: fixed sequences of steps (“run this command, then that”), brittle if the question changes mid-flight. The agent is not the primary interface — the process is. PMs doing discovery, competitive research, or feature validation need structured, evidence-backed outputs, but today they either click through rigid flows or drown in raw data without a partner that plans with them, explains coverage and confidence, and adapts when they say “go deeper on Reddit” or “skip Quora.”

Desired outcome: A chat-first research product where the agent is the main interface: the user states intent (e.g. “Research Notion’s iOS pain points for freelancers”), the system asks clarifying questions, proposes a research plan (sources, depth), waits for approval or edits, then executes while keeping the user in the loop — progress updates, mid-run steering, and evidence-backed insights that can be exported into a doc (PM-ready artifacts, not dumps). Under the hood, specialized sub-agents handle app/play store reviews, Reddit/forums, web/deep discovery, and analysis (sentiment, clustering, pain points, JTBD) with logic driven by coverage and confidence, not a single hard-coded brittle pipeline.

User

Primary: Product managers doing discovery, competitive research, or feature validation who want a research partner, not a batch script.

Secondary: Founders and UX researchers who need fast, structured evidence from public signals and internal notes (when wired in later).

Why This Problem Matters

Outcome before features: The north star is experiential — does this feel like a smart research partner? If the product still feels like a rigid automation, it fails the positioning. Shipping small, controllable experiments that prove planning + transparency + steering beats one-shot pipelines reduces wasted build and aligns with evidence over opinion (principles: structured thinking, user-problem-first).

Opportunity

Own the “agent as interface” slice for PM research: one controllable chat that unifies planning, tool orchestration, evidence gathering, and synthesis — differentiated from static research tools and from generic chat that does not show sources, coverage, or steerability.

Hypothesis

If we give PMs a conversational research agent that co-plans (clarify → propose plan → confirm), executes via specialized sub-agents with visible progress and mid-run steering, and returns citation-ready findings with export — then they will complete higher-quality discovery cycles faster and trust the output more than with a fixed multi-step wizard or raw LLM chat.

Risks / Open Questions

  • Source/API constraints: App stores, Reddit, forums, and web scraping may have ToS, rate limits, and reliability variance; need a clear legal/ethical stance and fallback behaviors.
  • Hallucination vs evidence: Synthesis must ground in retrieved artifacts; confidence and gaps must be explicit in the UX.
  • Scope creep: “Internal notes” and enterprise integrations can balloon scope — keep MVP to public web + export unless exploration proves otherwise.
  • Platform: Implementation app path in this monorepo is TBD (app: TBD); /explore should narrow MVP surface and stack.

Non-goals (initial)

  • Replacing the AI Product OS command pipeline for repo work (this issue describes a product to build, not a rewrite of /create-issue…/learning here).
  • Fully autonomous research with no human confirmation before expensive runs.
  • Guaranteed coverage of every possible source — agent must negotiate depth vs time with the user.

Success metrics (draft for /metric-plan / exploration)

  • North star (qualitative + proxy): User-reported “feels like a research partner” score or session rating; paired with task completion (plan approved → run finished → export used).
  • Supporting: Time to first approved plan, steering events per session (healthy adaptation), export usage, repeat sessions within 7 days.

Theme map (Linear-ready — refine in /create-plan)

ID Theme Notes
T0 Chat UX + plan contract Clarifying Qs, plan preview, approve/edit, cancel, progress stream
T1 Orchestration + sub-agents Store/Play, Reddit/forums, web discovery, analysis agents; routing
T2 Evidence + export Citations, clustering/JTBD outputs, doc export, gap/confidence surfacing

Next step

Send to Research Agent via /explore issue-013 to validate problem, competitors, MVP slice, and risks before /create-plan.