A source-grounded, two-stage pipeline for writing rigorous long-form academic work with AI.
Scriptorium is not a "write my essay" button. It is a method — a small set of carefully engineered prompt templates that split academic writing into the two tasks that current AI models each do well, and keep a human in control between them and after them.
The result is a long-form text (a coursework paper, term paper, thesis, dissertation chapter, literature review, or research report) that is structured, formatted to a defined standard, and — most importantly — grounded in real, traceable sources rather than fabricated citations.
A single large language model asked to "write my thesis" will produce fluent text and invented references. It conflates two jobs that should be separate:
- Finding and verifying what is actually known — locating authoritative sources, extracting their claims, numbers, and figures.
- Composing a coherent, well-styled argument from that material.
Scriptorium separates them.
┌─────────────┐ ┌──────────────────┐ ┌──────────────────┐
│ STAGE 0 │ │ STAGE 1 │ │ STAGE 2 │
│ Outline │ ───▶ │ Deep Research │ ───▶ │ Writing │
│ │ │ │ │ │
│ define the │ │ collect sourced │ │ compose the │
│ structure │ │ theses & data │ │ full text │
│ │ │ per section │ │ to the outline │
└─────────────┘ └──────────────────┘ └──────────────────┘
any model Google Gemini Claude
(optional) (Deep Research mode) (Project + documents)
│ │ │
└───── human review ──┴──── human review ─────────┘
- Stage 0 — Outline. Generate a complete table of contents for the work from its topic and requirements. Optional if your structure is already fixed by a supervisor or a template.
- Stage 1 — Deep Research. Run one structured research prompt per chapter in Google Gemini with Deep Research enabled. Each run returns detailed theses — what each author claims, the concrete data they cite — every one tied to a specific, checkable source. Gemini Deep Research is used here because it browses, reads, and aggregates many sources in one pass.
- Stage 2 — Writing. Load the collected research (plus any supporting documents) into a Claude Project and run one comprehensive writing prompt. Claude composes the full text against the outline, embeds citations from the research, and formats everything to your institution's standard.
A human reviews the output of every stage. Sources are verified by a person — AI fabricates citations, and Scriptorium's prompts are built to reduce that, not to eliminate the need to check.
- Separation of concerns. Research quality and prose quality are optimised independently. Each prompt asks one model to do one thing well.
- Traceability. Every claim in the final text descends from a thesis that names a real source. Nothing enters the writing stage unsourced.
- Structure first. The outline is fixed before research begins, so research effort maps cleanly onto sections and nothing is over- or under-covered.
- Style control. The writing prompt carries an explicit anti-pattern guide (see
docs/style-guide.md) so the prose does not read like generic machine output. - Human-in-the-loop by design. Review gates between stages are part of the method, not an afterthought.
| Path | Purpose |
|---|---|
templates/00-outline-prompt.md |
Stage 0 — generate the table of contents for the work. |
templates/01-deep-research-prompt.md |
Stage 1 — collect sourced theses for one chapter. Duplicate per chapter. |
templates/02-writing-prompt.md |
Stage 2 — compose the full, formatted text in Claude. |
docs/workflow.md |
The end-to-end process, step by step. |
docs/customization.md |
Every placeholder explained, with guidance on filling it. |
docs/style-guide.md |
The anti-AI-detection writing guide referenced by the writing prompt. |
examples/ |
A worked example on a neutral topic, showing filled-in placeholders. |
Every template is a Markdown file full of {{PLACEHOLDER}} tokens and bracketed [guidance notes]. You make a copy, replace the placeholders with your work's specifics, delete the notes, and paste the result into the model.
- Read
docs/workflow.md. It is short and walks the whole process. - Stage 0. Copy
templates/00-outline-prompt.md, fill in your topic and requirements, run it, and refine the resulting outline until you are satisfied. - Stage 1. For each chapter of that outline, copy
templates/01-deep-research-prompt.md, fill it in, and run it in Gemini Deep Research. Save each result to its own document. - Review. Spot-check the sources. Delete any thesis whose source you cannot verify.
- Stage 2. Create a Claude Project, upload the research documents and any supporting files, fill in
templates/02-writing-prompt.md, and run it — either in one pass or section by section. - Review and finish. Verify citations, fact-check key numbers, edit, and format.
docs/customization.md explains every field you need to fill.
Scriptorium is deliberately model- and topic-agnostic:
- Any field. The templates carry no subject matter of their own — you supply the field, topic, and outline.
- Any work type. Coursework, term paper, thesis, dissertation, literature review, research report.
- Any language. Set the output language in the placeholders. The structural style guidance is language-neutral; the lexical anti-patterns ship with English examples and a note on adapting them.
- Any standard. Citation style, length targets, and document formatting are all placeholders.
- Other models. Gemini Deep Research and Claude are the defaults the templates were tuned against; the prompts work with any comparable browsing-research tool and any capable long-context writer, with minor wording changes.
Scriptorium is a research and drafting aid. It does not replace your judgement, your understanding of the material, or your responsibility for the final work.
- Verify every source. AI models invent plausible-looking citations. A source that cannot be located and confirmed must be removed. This is not optional.
- Know your institution's rules. Policies on AI assistance in academic work vary widely and are changing fast. Using this pipeline may be permitted, restricted, or prohibited where you study. Find out before you use it.
- Disclose AI use where required. Many institutions require a statement of which AI tools were used and how. The writing prompt includes a dedicated section for exactly this — fill it in honestly.
- The work must be yours. Understand every argument, check every figure, and be able to defend the text. A pipeline that produces words you cannot stand behind has produced nothing useful.
The maintainers provide this tool for legitimate research assistance and accept no responsibility for misuse.
If Scriptorium saved you time or made your work better, please ⭐ star the repository — it takes a second and helps other students and researchers find it.
Issues and pull requests are welcome — particularly improvements to the prompt templates, additional language adaptations of the style guide, and worked examples in new fields. Keep examples on neutral, generic topics; do not commit real coursework.
MIT — see LICENSE.