Kiroku: Turning Paper Writers Into Advisors Using Multi-Agent Research Assistants
Hook
What if you stopped being the writer and became the advisor instead? Kiroku flips academic writing on its head by making AI agents your graduate students—researching, drafting, and revising while you provide strategic direction.
Context
Academic and technical writing follows a frustratingly linear process: stare at blank page, write terrible first draft, realize structure is wrong, repeat. The creator of Kiroku experienced this during PhD work at Stanford, going through formal advisor-student review cycles to produce papers and eventually a thesis. The insight was elegant: the advisor doesn’t write—they provide hypotheses, structural guidance, and iterative feedback while the student handles research, drafting, and revisions.
This role reversal isn’t just philosophical. Recent research suggests that LLMs help writers tackle complex topics through discussion rather than generation. Tools like Notion AI and Google Docs offer inline suggestions, but they keep you in the writer’s seat, staring at sentences. Kiroku delegates the entire drafting pipeline to a multi-agent system, letting you focus on what advisors actually do: define the hypothesis, approve section structure, and guide revisions until the message lands correctly.
Technical Insight
Kiroku’s architecture appears to separate concerns across planning, research, writing, and reflection phases—mirroring an advisor-student workflow. You start by defining a YAML configuration that specifies your hypothesis, document structure, and constraints:
title: "Writing Masterpieces when You Become the Advisor"
suggest_title: True
generate_citations: True
type_of_document: "research seminal paper"
area_of_paper: "AI and Computer Science"
section_names:
- Introduction
- Related Work
- Architecture
- Results
- Conclusions
number_of_paragraphs:
"Introduction": 4
"Related Work": 7
"Architecture": 4
hypothesis: "We want to show that turning writers into 'advisors' and multi-agent systems into 'advisees' enables interactive refinement from draft to publication."
instructions: "In the Introduction, first paragraph should discuss how ChatGPT changed the world."
references:
- "https://example.com/relevant-paper"
number_of_queries: 5
max_revisions: 3
temperature: 0.1This YAML acts as your research proposal. The hypothesis field defines the core argument you want to make, while instructions let you append specific guidance like “The first paragraph of Introduction should discuss the revolution created with ChatGPT.” These instructions get appended to the hypothesis, informing downstream agent decisions.
The multi-agent pipeline kicks off with research planning. Kiroku generates search queries based on your hypothesis and feeds them to Tavily’s API—grounding the writing in actual research rather than hallucinated citations. If you set number_of_queries: 5, the system formulates five targeted questions to gather evidence supporting your hypothesis.
Next appears to come topic sentence generation. For each section defined in section_names, Kiroku creates topic sentences that map to your specified number_of_paragraphs. If you want four paragraphs in the Introduction, you get four topic sentences that scaffold the section’s logical flow. This mirrors how advisors push students to outline before writing.
The expansion phase appears to turn topic sentences into full paragraphs using OpenAI’s LLM, with the temperature parameter typically set to low values like 0.1 for technical accuracy. If generate_citations: True, the system weaves in references from the Tavily research phase, producing academic-style citations rather than generic statements.
The reflection mechanism is where Kiroku’s advisor-student dynamic shines. The max_revisions parameter controls how many times agents perform self-critique and rewriting. Setting this to 3 means three cycles of reflection and revision—the agent reviews its own output against your hypothesis and instructions, identifies weaknesses, and rewrites. This iterative loop mimics advisor feedback sessions where drafts progressively improve.
Kiroku includes a Gradio web interface for interaction, though documentation candidly mentions image handling requires special /file= syntax (suggesting UI maturity issues). The system uses PANDOC for document format conversion, supporting outputs like PDF, DOCX, and LaTeX—critical for academic workflows where final formatting matters.
The original codebase drew inspiration from a LangGraph course by Harrison Chase and Rotem Weiss, but the creator notes that even the prompts have diverged significantly from the original course material.
Gotcha
Kiroku demands both OpenAI and Tavily API keys, with no support for local or open-source LLMs. If you’re writing a thesis with multiple revision cycles across many paragraphs, API costs accumulate. The reliance on external services also means no offline operation—exactly when you might want focus mode during writing sessions.
Python version support is between 3.7 and 3.11, and the external PANDOC dependency creates installation friction. The README instructs macOS users to brew instal pandoc (with a typo in the command), and supports Windows through WSL. The Gradio interface limitation requiring /file= syntax for images suggests rough edges—the documentation literally includes a section titled ‘Bugs :-)’ acknowledging known issues. This isn’t production-ready software; it’s a research tool that expects technical users comfortable debugging.
The YAML-driven workflow assumes you have a clear hypothesis and section structure before starting. If you’re in exploratory mode—figuring out what you want to say by writing—Kiroku’s structured approach feels constraining. You can’t just free-write a paragraph and see where it goes; you must define sections, paragraph counts, and hypotheses upfront. This rigidity is intentional (it mirrors academic writing norms), but it makes Kiroku unsuitable for creative or exploratory content.
Verdict
Use Kiroku if you’re writing structured academic or technical documents—research papers, theses, technical reports, grant proposals—where you have a clear hypothesis and know the section outline but dread the research synthesis and drafting grind. It excels when you embrace the advisor role: defining strategic direction, providing iterative feedback, and refining arguments while agents handle the tedious paragraph construction. PhD students and researchers familiar with advisor-student dynamics will recognize the workflow immediately. Skip it if you need freeform creative writing, can’t afford ongoing API costs, require local-only solutions for privacy, or prefer direct control over every sentence. The tool’s value proposition is delegation—if you’re not willing to let agents draft while you guide, you’ll fight the system instead of leveraging it. Also skip if you’re writing marketing copy, blog posts, or narrative content; Kiroku’s citation-heavy, hypothesis-driven approach is built for academic rigor, not viral engagement.
