research-planning

name	research-planning
description	Create, manage, and iterate on research plan documents in the Obsidian vault. Use when the user asks for a research plan, literature review roadmap, or when you need to formalize your own working research strategy.
tools	["create_plan","list_plans","get_plan","update_plan","delete_plan"]
tags	["planning","research","workflow"]

Research Planning

Create and manage structured markdown plan documents directly in the Obsidian vault. Plans are automatically indexed into the RAG system so you can reference and search them alongside your research papers.

Two Types of Plans

Internal Plans (plan_type="internal")

Stored in thoth/_thoth/plans/ — your own working research strategy. Use these to:

• Map out a multi-step research approach before starting
• Track which questions remain open
• Note decisions and why you made them
• Draft a plan the user can review and edit

User Plans (plan_type="user")

Stored in thoth/plans/ — formalized deliverables for the user. Use these when:

• The user asks "can you make me a research plan for X"
• You want to present a structured roadmap the user can follow independently
• You're summarizing a completed research arc into a reusable plan

Recommended Plan Structure

Always include these sections in the content argument:

## Objective
One paragraph: what research question are we answering and why.

## Steps
1. Step one (specific, actionable)
2. Step two
3. ...

## Key Papers / Sources to Check
- Any known papers or sources worth starting from

## Open Questions
- Unknowns to resolve during the research

## Notes
Running notes as the plan evolves.

Workflow

Starting a new research session

1. Call create_plan with plan_type="internal" to write out your approach
2. Proceed with discovery and analysis
3. Call update_plan as your understanding evolves (update status, add notes)
4. When ready to present to the user, call create_plan with plan_type="user" with the polished version

When the user explicitly asks for a plan

1. Clarify scope if needed
2. Call create_plan with plan_type="user" and a thorough content
3. Tell the user the plan is saved in their vault at thoth/plans/{plan_id}.md

Checking existing plans

list_plans                          # all plans
list_plans(plan_type="internal")    # only your working plans
list_plans(status="active")         # only active plans
get_plan(plan_id="my-plan-slug")    # read full content

Updating a plan

update_plan(
    plan_id="my-plan-slug",
    status="complete",              # or "draft", "active", "archived"
    content="## Objective\n..."     # replaces full body
)

Plan IDs

Plan IDs are auto-generated slugs from the title. For example:

• "Literature Review: Transformers 2026" → literature-review-transformers-2026
• "My Research Strategy" → my-research-strategy

Use list_plans if you need to look up an ID.

RAG Integration

Plans are indexed into the same vector store as your research papers. This means:

• answer_research_question will surface relevant plan content in its context
• Searching for a topic will return matching plans alongside papers
• You can refer to plans when answering questions about research strategy

Status Values

Status	Meaning
draft	Not ready yet, still being drafted
active	In progress or currently relevant
complete	Research arc finished
archived	No longer active but kept for reference

Example

User: "Can you create a research plan for understanding how attention mechanisms work in LLMs?"

→ create_plan(
    title="Research Plan: Attention Mechanisms in LLMs",
    plan_type="user",
    content="""
## Objective
Understand how attention mechanisms work in large language models,
from the original transformer paper through modern variants.

## Steps
1. Read Vaswani et al. 2017 (Attention Is All You Need)
2. Search for survey papers on attention variants
3. Compare: MHA, MQA, GQA, FlashAttention architectures
4. Identify efficiency vs. quality trade-offs
5. Summarise findings

## Key Papers / Sources to Check
- Vaswani et al. 2017
- Flash Attention (Dao et al.)
- GQA paper (Ainslie et al.)

## Open Questions
- How does RoPE interact with attention?
- What are the memory bottlenecks?

## Notes
Starting fresh — no prior papers in collection on this topic.
""",
    tags=["attention", "transformers", "llm-architecture"]
)