clone-research

Clone Research

Target: $ARGUMENTS

Agent Strategy

Spawn 4 parallel research agents. Each produces 1-3 documents from external research.

All agents: subagent_type: general-purpose, run_in_background: true

References: agent-roles

Execution

1. Parse Input

Target: $ARGUMENTS

If the target above is non-empty, use it immediately — do NOT ask the user to confirm or re-provide it. Parse it as follows:

• URL (starts with http): use as-is for WebFetch, extract company name from domain
• Company name (no URL): construct likely URLs (https://{name}.com, https://www.{name}.com)

If the target above is empty, ask the user what product to research and wait for their response.

Store:

• COMPANY_NAME: Human-readable name (e.g., "Linear")
• PRIMARY_URL: Main product URL (e.g., "https://linear.app")
• SLUG: kebab-case for directory (e.g., "linear")

Create output directory: clone-research/{SLUG}/

IMPORTANT: When a target is provided, begin Phase 2 immediately after parsing. Do not pause for user input.

References: workflow

2. Spawn Agents in Parallel

Spawn all 4 agents in ONE parallel Task tool call. Each agent receives:

• Company name and URL
• Their specific mandate and prompt template from agent-roles reference
• The relevant document templates from output-templates reference
• Document standards (naming, feature IDs, entity naming, confidence levels, citations)
• Research techniques for their specific role

Build each agent's prompt by:

1. Taking the prompt template from the agent-roles reference
2. Replacing {COMPANY_NAME}, {PRIMARY_URL}, {OUTPUT_DIR} with parsed values
3. Pasting the document templates they own from the output-templates reference
4. Including document standards and their research techniques

References: agent-roles, output-templates, document-standards, research-techniques

3. Cross-Reference & Index

After all agents complete:

1. Read all 9 documents from clone-research/{SLUG}/
2. Write clone-research/{SLUG}/00-INDEX.md with executive summary, reading order, and cross-reference map
3. Verify feature IDs (F1, F2...) and entity names are consistent across documents
4. Note any inconsistencies in the INDEX under "Known Gaps"
5. Write AGENTS.md at the project root (current working directory) using the AGENTS template — includes operating instructions, domain model, architecture stack, and phased work plan (Phase 0 through 4) with per-phase checklists; this file is the primary context document for all future agents building the clone
6. Run ln -s AGENTS.md CLAUDE.md at the project root — creates a relative symlink so Claude Code auto-loads the context when opening the project; only ever edit AGENTS.md going forward

References: workflow, output-templates

Output

9 research documents + index in clone-research/{SLUG}/, plus AGENTS.md and CLAUDE.md symlink at the project root:

./                               # Project root (current working directory)
  AGENTS.md                      # Project context: mission, scope, stack, work phases
  CLAUDE.md -> AGENTS.md         # Symlink — Claude Code auto-loads this; edit AGENTS.md only
  clone-research/{slug}/
    00-INDEX.md                  # Executive summary + reading order
    01-product-thesis.md         # North star, market, opportunity
    02-feature-priority-matrix.md  # Feature inventory P0-P3
    03-mvp-scope-contract.md     # In/out scope, roadmap phases
    04-core-user-journeys.md     # Personas, JTBD, step-by-step flows
    05-domain-model.md           # Entities, relationships, lifecycles
    06-system-architecture.md    # Observed + recommended stack
    07-revenue-pricing-model.md  # Tiers, gating, upgrade triggers
    08-api-surface-spec.md       # Endpoints, pagination, errors
    09-design-system-brief.md    # Colors, type, spacing, components

Present to user: file list, executive summary, gaps, and suggested reading order.

Anti-Patterns

• Don't use Explore agents: Sub-agents need WebSearch and WebFetch for external research. Use general-purpose only.
• Don't collapse agents: Each agent has a distinct research lens. Combining them loses depth.
• Don't fabricate data: If information isn't found, say "Not publicly available" rather than guessing.
• Don't skip citations: Every factual claim must reference a source URL.
• Don't run agents sequentially: All 4 agents are independent — spawn them in parallel.
• Don't skip the INDEX: The cross-reference map is essential for downstream consumers.

Example Invocations

/clone-research https://linear.app
/clone-research Notion
/clone-research https://www.figma.com
/clone-research Vercel
/clone-research Superhuman

Notes

• Total runtime: typically 4-10 minutes depending on the target's web presence.
• All 4 agents run in background for maximum parallelism.
• If an agent fails or returns thin results, note the gap in the INDEX rather than blocking.
• For private/stealth companies with minimal web presence, agents will produce thinner reports — this is expected.
• Documents are consumed by builder agents downstream — consistency in naming, feature IDs, and entity names matters.