Install from the repository
The skill package includes instructions, a deterministic retrieval script, and the released catalog. Core search works locally and does not require manuscript upload. Review the repository and permissions before installing any agent extension.
$skill-installer https://github.com/rezaprama/RhetoriLex/tree/main/skills/rhetorilexYou can also clone the repository and point your agent to skills/rhetorilex/SKILL.md. The Python command-line interface supports the same catalog for reproducible searches.
Use the installable web app
The Paraphrase Workbench can run as a browser app. It works without an AI endpoint by generating a safe prompt, or it can call an OpenAI-compatible or Gemini-compatible endpoint that the user enters. Tokens are handled in the browser and are stored only when the user chooses to remember settings.
The hosted static page applies a three-call daily browser guard for endpoint calls. This is not a substitute for server-side security. A shared public model key should use a backend with server-held secrets, daily quota storage, abuse protection, and manuscript logging disabled by default.
Ask for purpose and constraints
A useful prompt names the rhetorical task, manuscript section, evidence type, and protected meaning. It also tells the agent what it must not infer. These examples are starting points:
Rewrite this observational result without causal language. Preserve the estimate, interval, population, and citation. Explain any wording that would require a causal design.
Paraphrase this sentence. Preserve citation, number, unit, negation, comparison direction, and uncertainty. Return a checklist showing each invariant.
Find three patterns for a cautious discussion interpretation. Prefer tentative or bounded claim strength and low risk. Show evidence requirements before the templates.
Read the output as a proposal
The agent should identify the move, return candidate patterns, explain claim strength, and warn when evidence does not meet a requirement. It should not invent results, references, methods, or consensus. A causal pattern is not permission to make a causal claim. The design and identifying assumptions must already justify it.
After adapting a pattern, compare the sentence with the manuscript evidence. Verify names, citations, numbers, units, direction, negation, modality, population, time, and limitations. Ask a subject specialist when interpretation depends on domain knowledge.
Keep searches reproducible
Record the catalog version or commit, query, filters, and selected pattern ID when the wording decision matters. The command-line interface can return stable IDs for this purpose. Search is deterministic, so the same catalog and query produce an auditable result order.
rhetorilex search "cautious interpretation" --limit 3
rhetorilex show RLX-DISCUSS-INTERPRET-001