# AgentTune Editable communication preferences for AI assistants. Original templates and code are MIT; questionnaire terms are instrument-specific. Templates are optional hypotheses, not diagnoses or demonstrated capability improvements. ## Discover - [Tuning catalog: 43 templates](https://agent-tune.com/library/index.json): canonical URLs, body-only exports, revisions and evidence status. - [Guide and research catalog](https://agent-tune.com/resources/catalog.json): focused resources and page-specific Markdown. - [Platform registry](https://agent-tune.com/resources/platforms.json): version 2026-09-25.1, app and destination instructions shared with the UI. - [Installation protocol](https://agent-tune.com/resources/install-protocol.md): authorization, preservation, verification and undo. - [Free tools](https://agent-tune.com/resources/tools/catalog.json): eight browser tools and a local text-processing API; discover with get_free_tools. - [Generator](https://agent-tune.com/tools/custom-instructions-generator): start with explicit preferences; no inferred personality type. - [Local deterministic scoring](https://agent-tune.com/resources/scoring/README.md): all five available adaptations, strict versioned inputs and incomplete-answer handling. - [Test specifications](https://agent-tune.com/tests/): current availability, scoring and limitations. - [MCP](https://agent-tune.com/mcp): list_tunings, get_tuning, get_test_spec, list_resources, get_resource and get_free_tools. Stateless Streamable HTTP; POST JSON-RPC, no authentication. ## Use Read only the resource needed for the user’s request. Fetching content does not authorize installation. Use body-only text in instructions and keep metadata separate. Preserve existing files, permissions, accuracy and material uncertainty. Explicit user preferences resolve conflicts; no personality framework has automatic priority. Research records questionnaire responses, not validated improvements in task performance. Current Big Five results report raw scores and item means, without population percentiles. Historical studies retain their original scoring protocol. Report ties and missing values. See the article-specific sources and methods before citing a result. # Installing AgentTune preferences Version: 2026-09-25.1 · [Platform registry](https://agent-tune.com/resources/platforms.json) · [Generator](https://agent-tune.com/tools/custom-instructions-generator) 1. A request to read a tuning is not permission to install it. Confirm the requested destination and scope. 2. Fetch the tuning body separately from metadata. Review the wording with the user; explicit preferences and the current task take precedence. No framework has automatic priority. Preserve accuracy, relevant uncertainty and permissions. 3. Choose the app and personal, project or conversation destination using the registry. Interface labels can vary by account. API clients use the provider’s supported instruction field; omit installation metadata and select the model separately. 4. Preserve existing instructions. Add one block bounded by `` and ``. When updating, replace only that block. If markers are duplicated or incomplete, stop and resolve the ambiguity. Preserve format-specific front matter. Do not overwrite an existing file with a downloaded template. 5. Confirm storage by rereading the exact saved text. This establishes that it was saved, not that it is always applied. 6. In a new conversation or task, try a factual question, a planning task, and a critique request without repeating the desired style. Compare observations to the selected preferences and note counterexamples. Do not use a greeting that dictates its own expected response as a compliance test. 7. To undo, remove only the managed block or exact added text; preserve all other content. If the block was subsequently edited, review the difference before removal. ## Evidence limits Personality templates are editorial communication suggestions, not diagnoses or proven performance improvements. Test scores describe responses to a questionnaire. Model self-reports do not establish an inner personality. See each research article for its data provenance and limitations. ## Compact exports The ChatGPT export uses a conservative 1,500-character budget. This is AgentTune’s export budget, not a claim about every current product field limit. The registry and generator share the same destination instructions. ## Data handling Score questionnaires locally. Do not send answers to AgentTune MCP; its tools only retrieve public resources. Report missing responses and ties explicitly. Never infer consent or a personality type from a default selection. ## Platform destinations ### Any chat Start a new chat with your chosen model and paste this prompt. Follow it with your task. For persistent preferences, use a supported saved-instructions destination. ### Muse · saved preferences Paste this request into Muse and review the proposed change. It asks Muse to update only “How to work with me” in Soul.md, preserving other preferences. Muse setup guide → ### Codex · personal instructions Add this section to ~/.codex/AGENTS.md for your personal defaults, or to a project’s AGENTS.md for shared project preferences. Preserve existing instructions, then start a new task. Codex setup guide → ### ChatGPT · project instructions Open your ChatGPT project settings and add this block to its instructions. Start a chat in that project with your chosen model. ### ChatGPT · custom instructions Open Settings → Personalization → Custom Instructions and add this compact block. This export uses a conservative 1,500-character budget; the controls available in your app may differ. Use project instructions or a new chat for the full tuning. ### Claude · account instructions Open Claude’s Settings → General → Instructions for Claude and add this block. Review it alongside existing preferences. Official personalization guide → ### Claude · project instructions Open a Claude project and add this block to its instructions. Choose Fable 5.1 or Opus 5.5 in Claude where available, then start a project chat. Claude setup guide → ### Claude Code · output style Save this as ~/.claude/output-styles/agenttune.md and select it with /output-style. If that file exists, preserve its header and other instructions; update only the marked AgentTune block. The header keeps Claude Code’s coding instructions enabled. Claude Code setup guide → ### Grok · chat or Bot Paste this in a new Grok chat, or add it to your Grok Bot’s Description alongside its job. Select Grok 4.7 wherever your app offers it; a prompt cannot switch models. Grok Bot setup guide → ### Cursor · project rule Save this in .cursor/rules/agenttune.mdc. Merge with any existing file. This always-on project rule can be shared with collaborators, so use preferences appropriate for the project. ### GitHub Copilot · repository instructions Append this section to .github/copilot-instructions.md, preserving existing repository instructions. Use preferences appropriate for everyone sharing this project. ### Gemini · Gem instructions Create or edit a Gem in Gemini and add this block to its instructions alongside the Gem’s job. Save it, then open a chat with that Gem. ### Gemini CLI · GEMINI.md Append this section to your project’s GEMINI.md or personal ~/.gemini/GEMINI.md. Preserve existing instructions and reload the context. ### Hermes · persona Add this block to your Hermes persona alongside its role. Preserve existing instructions and select the model separately. ### OpenClaw · AGENTS.md Add this section to the workspace’s AGENTS.md. Keep existing identity, permissions, and task instructions. ### API · instruction text Send this text using OpenAI Responses instructions, Anthropic’s system field, or a Grok system message. Set the model ID separately. Include the instructions on each request where needed; this is prompt text, not a complete API request. ### MCP · connect AgentTune Run this setup command in Claude Code. For other MCP clients, add https://agent-tune.com/mcp as a remote server. Then ask your agent to find and apply your tuning. This copies a connector command, not your selected tuning.