Research Idea Explorer

• formatting an existing result
• inspecting graph, summary, or neighbors
• writing feedback
• explaining the output schema or command usage

The workflow is:

1. Start with a literature-guided single-pass divergence across multiple research moves. The built-in persona set remains an internal way to keep those moves orthogonal, but it should not dominate the user-facing framing.
2. Convert only the strongest brainstorm seeds into the shared internal schema: Object, Puzzle, Claim, Contrast, Evidence, Scope, Stakes.
3. Reject ideas that are obvious near-duplicates or amount to a thin "method + topic" pairing.
4. If the user asks to deepen, branch, compare, or continue from an accepted direction, run a second literature-guided mutation round around the retained frontier families.
5. Push each retained idea one step further by adding: data source, minimal runnable design, likely failure mode, and why the question matters now.
6. When available, use the persistent memory graph so the session does not keep revisiting the same research neighborhoods.

Preference handling

When the user expresses stable preferences in conversation, carry them into the CLI call instead of treating them as disposable chat context.

Typical examples:

• preferred direction: prefer causal identification, focus on policy relevance
• constraints: avoid survey, don't do LLM, less benchmark work

Default behavior:

• Pass those signals through --preference-note.
• Use --remember-preferences topic when the preference is specific to the current topic.
• Use --remember-preferences global only when the user clearly wants the same preference to apply across topics.
• Reuse the same memory path so stored preferences can affect the next ideas run.

CLI mapping

When you want the packaged command to do the work instead of manually re-creating the flow:

• generate ideas: research-idea-explorer ideas --query "..."
• inspect memory: research-idea-explorer graph --memory ./data/memory/cli-memory.json
• list ideas for feedback: research-idea-explorer feedback --memory ./data/memory/cli-memory.json
• accept or reject an idea: research-idea-explorer feedback --memory ./data/memory/cli-memory.json --idea-id <idea-id> --decision accepted
• generate ideas while storing the user's conversation preferences: research-idea-explorer ideas --query "..." --preference-note "prefer causal identification, avoid survey" --remember-preferences topic
• store preferences during feedback on a retained idea: research-idea-explorer feedback --memory ./data/memory/cli-memory.json --idea-id <idea-id> --decision accepted --note "prefer causal identification, avoid survey" --remember-preferences topic
• inspect stored preference profiles: research-idea-explorer graph --memory ./data/memory/cli-memory.json --view preferences

If only the repo-local version is available, the same commands can be run with python -m research_idea_explorer.cli.

Guardrails

• Do not output only titles. Output brainstorm seeds first when useful, then concise research cards.
• Do not overuse Blind spot if a stronger puzzle such as Conflict, Distortion, or Mechanism unknown is available.
• Treat Scale and Counterfactual as part of contrast and scope unless the user explicitly wants them foregrounded.
• Search literature by default before proposing ideas, and use metadata to identify crowded areas before proposing ideas.
• Prefer diversity of research moves over superficial variation in wording.

Diagram provenance

If you produce a diagram, distinguish clearly between:

• a computed graph generated from stored nodes and edges
• a literature-derived map synthesized from retrieved papers
• a conceptual sketch synthesized by the agent

Do not present a hand-authored conceptual diagram as an automatically generated network graph.

Output shape

Each default research card should contain:

• Title
• Abstract
• Design
• Distinctiveness
• Significance

The richer internal schema can still be used when deeper analysis is needed, but the default user-facing card should stay concise.

References

Read references/schema.md when you need the field meanings, internal move set, or puzzle definitions. Read references/retrieval.md when literature retrieval is available and you need to decide between lexical, embedding, or graph expansion.