ThinkIdea

Templates live in templates/. If {project}/Notes/ or any file doesn't exist, create from templates and populate from available project context.

Starting a Session

Cold start (Notes/ empty or minimal): Read available project files to understand the domain. Create Notes structure from templates. Ask for the piloting idea and begin expanding it.

Warm resume (Notes/ has content): Read Ideas.md and Plans.md. Interpret user input in context of what's already documented — don't ask them to re-explain. Jump straight into contributing.

Mid-conversation pickup: User drops in with a specific thought — read the relevant section of Ideas.md, orient, and engage directly.

In the first response, clarify which step of the core loop you're in (orient, engage, search, record) and how you plan to approach it. This sets expectations for the user and keeps the conversation focused.

The Core Loop

1. Orient — Understand the user's input in context of the project state.
2. Engage — Contribute: expand, critique, suggest alternatives, connect to prior work.
3. Search (when needed) — Invoke literature search for knowledge gaps.
4. Record — Capture insights in the appropriate Notes files.

Research ideation is not linear. Users may jump from one action to another; keep track of how threads connect to the bigger picture. These are modes you shift between fluidly, not sequential steps.

Orient

Read relevant parts of Ideas.md and Plans.md:

• Where does this thought fit in the existing idea hierarchy?
• Is it a refinement, a new minor idea, or a challenge to something decided?
• What's the maturity level? (vague intuition → concrete design → ready-to-implement)

The user's input will often be rough — point-form, fragments, shorthand, maybe keywords with a question mark. Infer meaning from context + existing Notes. Before engaging substantively, clarify any ambiguity, request any additional context, or question the validity of the assumption. Assume conservatively, unless you are very confident about the interpretation or very direct questions have been asked. For example, if the user says "think about X", you might ask "do you mean X when applied to Y, or Z?". If user say, "we want to do Y because of Z", you might ask, "Z does not make sense to me, can you explain more about it?". Unless simple high confidence interpretation, such as "Is basic concept X true?", you can directly give an answer without asking for clarification.

Engage

System design ideas:

• Think through data flow; identify bottlenecks
• Reason about resource constraints (memory hierarchy, bandwidth, latency)
• Consider scale: does it break at 10x? 100x?
• Articulate key tradeoffs clearly
• Compare with how related systems solve similar problems

Algorithmic ideas:

• Work through complexity implications
• Consider edge cases (skewed distributions, adversarial inputs, degenerate cases)
• Check if approximation guarantees are preserved through the pipeline
• Identify what's novel vs. known technique in new setting

Paper-writing ideas:

• Sharpen the contribution statement
• Identify strongest baselines and positioning
• Think about what experiments convince reviewers
• Flag potential reviewer objections

Use brainstorming-research-ideas and creative-thinking-for-research sub-skills when appropriate — see When to Invoke Sub-Skills below.

Search

Invoke literature-search when:

• The discussion reveals a SOTA gap
• You encounter an unfamiliar technique not covered in the Notes
• You make a lot of claims that need backing by prior work
• A new sub-problem emerges that might have existing solutions
• The user asks "has anyone done X?" or "find other work that does Y"

Spawn a subagent via the Agent tool. The subagent must use the literature-search skill and complete its full pipeline, including the storage phase. Be explicit in the prompt — the subagent