Skip to content

Best practices ​

Effective prompting ​

Specific instructions reduce ambiguity so the model spends its tokens on your actual goal rather than guessing scope or style.

Name the target sections and goal ​

Instead of "Make this better", specify the section and what to tighten:

Tighten the methodology section (\section{Variational Ansatz}): remove hedging and redundant transition phrases, and keep every symbol definition anchored before its first use.

State what to preserve ​

Explicitly protect macros, labels, theorem numbering, and settled sections:

Revise the introduction and discussion for clarity. Preserve all \cite{} keys, equation environments, \label{} tags, and custom macros. Do not modify the results section.

Separate content derivations from prose polishing ​

Run derivation, proof, or literature tasks (research, prover, lean, or the Orchestrator) first. Once the technical claims and equations are settled, ask for a polish or a proofread of the finished .tex file (the edit program).

Model selection ​

Match model capability to task complexity: overpowered models waste money, underpowered ones produce poor results.

Match model to taskcapability vs. complexity
SimpleCorrections, quick edits
openai/gpt-6-lunadeepseek/deepseek-flash@noneanthropic/claude-haiku-5-5@none
ComplexTransformations, rewrites
anthropic/claude-fable-5-1anthropic/claude-opus-5-5openai/gpt-6.1-solgoogle/gemini-4-argon
Reasoning-heavyDeep, multi-step thinking
anthropic/claude-fable-5-1anthropic/claude-sonnet-5-5anthropic/claude-opus-5-5deepseek/deepseek-flash

Find the row that matches your task, then pick one of the recommended model handles.

Read the AI models for detailed comparisons.

Project and LaTeX organization ​

project/
├── AGENTS.md
├── main.tex
├── chapters/
├── figures/
├── refs.bib
└── .texra/
  • Keep an AGENTS.md at the workspace root: record project-wide notation conventions, custom macros, compilation commands, and forbidden edits so every agent session inherits them automatically.
  • Give a program only the files it needs: the target .tex file as input and shared preamble or .bib files as context.
  • Format before running latexdiff: run latexindent or tex-fmt before comparing revisions so line-wrapping differences do not obscure prose edits.
  • Archive milestone revisions: use Archive outputs to History in the Tasks view menu to copy key revisions into History/, and delete stale run folders when no longer needed.

Workflow integration ​

Staged approach ​

A typical project moves through five stages, each owned by one or more agents or programs:

Staged workflowresearch → finalization
  1. ResearchFind papers and verify ideas
    searchresearch
  2. WritingDraft LaTeX content interactively
    researchorchestrator
  3. DevelopmentRefine clarity and style
    edit
  4. VisualizationProduce figures and slides
    researchpresenter
  5. FinalizationProofread the finished draft
    edit

A typical project flows from research to finalization, with each stage handed to the agent or program best suited for it.

Quality assurance ​

Review AI-generated content after every run, using the same checks each time:

Review every AI-generated changeafter each run
Compiles without LaTeX errors
Cross-references & citations resolve
Figure & table numbering correct
Math expressions verified
No omissions or duplications

A short review pass after each run catches the errors automation cannot see for itself.

Use LaTeX Diff to visualize changes between versions.

Next steps ​