Elysia Tools
Navigation
Workflow Playbook
Prepare LLM inputs by structuring prompts, checking token budgets, translating or identifying language, cleaning PDF evidence, and reviewing injection, math, regex, and data risks before use.
Hubs
A useful LLM call is rarely just one sentence. It usually combines an instruction, source evidence, language expectations, data format rules, and risk notes. This workflow keeps those parts separate until they are ready to be assembled.
Start with prompt-optimizer to make the role, task, constraints, and examples visible. Then use ai-token-estimator to check whether the planned instruction and evidence fit the target context window. The estimate is planning evidence, not a promise that the model will answer in a specific way.
If a PDF will become grounding text, extract a cleaner text version with pdf-to-clean-text-for-llm and run pdf-prompt-injection-scanner before the content reaches an automation or chat pipeline. Hidden, tiny, off-page, or layered text should be reviewed as risk evidence rather than silently folded into the prompt.
For multilingual input, combine prompt-translator with ai-language-detector so the final instruction names the intended language and preserves important terminology. If the prompt references tables, JSON, CSV, equations, or regular expressions, use ai-data-normalizer, ai-math-solver, and ai-regex-explainer to make those details explicit.
The final deliverable should be a prepared input bundle: the prompt, token notes, cleaned source text, language assumptions, normalized data, and any safety warnings. It should also state what remains uncertain. These tools can improve clarity and reviewability, but they do not guarantee the quality, accuracy, or safety of the model's eventual output.
Workflow playbook
Rewrite the rough instruction into a reviewable structure, then estimate token usage before adding source material or examples.
Translate prompts for multilingual use and detect source-language mix so terminology, locale, and script assumptions are visible.
Extract clean PDF text for grounding and scan hidden, off-page, or tiny text that could contaminate an LLM input.
Clean messy structured data and make math problems or regular expressions explicit before they are referenced by the final prompt.
Start with prompt optimization and token estimation for unclear instructions, PDF cleanup and injection scanning for document evidence, or data and language checks when the input is multilingual or structured.
Use the math and regex helpers to make equations or patterns explicit before they are embedded in a prompt. These checks improve reviewability but do not guarantee the model's final answer.