PM DeepResearch

PM DeepResearch is a product manager's deep competitive-research skill for Claude Code. It turns a product decision ("should we enter / differentiate / build / upgrade to AI?") into a decision-oriented, evidence-complete 13-chapter competitive report, applying a five-dimension product methodology (Job & real competitive set, capability matrix, Kano, ODI opportunity gaps, positioning & whitespace) with explicit epistemic tagging and falsifiability.

It is a Skill / orchestration layer (Layer 1) that runs on top of the Lapis MCP research core (Rust, upstream 4o3F/Lapis, AGPL-3.0). Lapis owns MCP execution, provider calls, agent loops, budget guards, schema validation, and byte-equal evidence provenance; PM DeepResearch carries the product methodology via prompt assets + Skill-layer assembly. Lapis is consumed unchanged — any engine-schema needs are filed upstream as requirements, not patched in.

Status: Phase 3 — v2.0 core validated. A 6-aspect Deep run on a golden topic (Strava AI-coaching upgrade) produced a 13-chapter report scoring 22/24 on the project rubric. Structure reorg (vendored-engine extraction + one-click installer) is scheduled for Phase 4.

The skill

Path	What it is
skills/pm-deep-research/SKILL.md	The competitive deep-research skill entry: decision-intent routing, complexity tiers, five-dim decomposition, persona assembly, Lapis MCP calls, evidence post-processing, gap audit + quality-floor self-verification, 13-chapter report.
skills/pm-deep-research/prompts/layer1/	Orchestration prompts: task-decomposition, agent-allocation, evidence-postprocess, final-report, claude-only-degradation.
skills/pm-deep-research/prompts/layer2/	Persona prompts: persona-strategist, persona-experience-analyst.
skills/deep-research.md	Generic Lapis research skill kept as a fallback / base.

When the Lapis MCP server is unavailable, the skill degrades to Claude-only (search MCP directly, same methodology) — degradation is not failure; the methodology lift is pure prompt capability.

Key documents

Document	Purpose
ROADMAP.md	Single source of truth for the plan (rolling-wave).
docs/specs/pm-deep-research-competitive-research-spec.md	Canonical competitive-research spec (five-dim, 13 chapters, evidence discipline).
docs/specs/orchestration-interface.md	Layer 1 ↔ Lapis interface (request/result, budget, evidence post-processing, degradation).
docs/evaluation/rubric.md	12-dimension scoring rubric + prose floor; golden-sample anchor.
docs/decisions/	Architecture decision records (ADR-0001…0006).

Vendored Lapis engine

This repository currently vendors the Lapis engine source (crates/, Cargo.toml, lapis.example.toml) so the MCP core can be built and run locally during development and validation. The engine is upstream-owned (4o3F/Lapis); we do not develop it here. Per ADR-0002, Phase 4 replaces the vendored copy with a one-click installer that fetches and builds upstream Lapis, at which point the vendored source is extracted.

Build and run the engine locally:

cargo build --release
cp lapis.example.toml lapis.toml          # then enable providers + export the referenced API keys
./target/release/lapis serve --config lapis.toml

The server speaks MCP over stdio and exposes two tools:

Tool	Purpose
aspect_research	Runs one research aspect and returns an AspectResearchResult.
deep_research	Runs multiple research aspects and returns a DeepResearchResult.

Supported request schema_version: 0.1. Engine-side details:

Document	Purpose
docs/mcp-usage.md	MCP client interface and tool schemas.
docs/configuration.md	Runtime configuration, providers, budgets, logging.
docs/development.md	Repository layout and contributor commands.

Requirements

• Rust toolchain with Rust 2024 edition support (to build the vendored engine).
• Claude Code (or another MCP client) able to run stdio MCP servers.
• At least one enabled model provider, and a search provider for aspects that allow search.

Common development checks

cargo fmt --all -- --check
cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings

License

GNU Affero General Public License v3.0. See LICENSE for details.