/clarify
Ask only the questions whose answers materially change cost, architecture, safety or user experience. Put the recommended answer first.
LLM field guide / 12 major products
Know what really works in each AI.
Slash commands are not universal. This reference names the exact product surface, links to first-party documentation and labels every entry as a literal control, documented workflow or portable prompt pattern.
Literal product controlDocumented workflowPortable prompt pattern
Maviter portable patterns / Any conversational LLM
Useful names for repeatable instructions.
These seven labels are Maviter shorthand, not secret platform commands. Paste the full instruction after the label, or save it in the platform’s supported instruction system.
/research
Identify time-sensitive claims, prefer primary sources and separate sourced fact from inference.
/architect
Convert the approved outcome into phases, dependencies, acceptance checks and freeze boundaries.
/protect
Inventory existing behaviour, data and infrastructure that the next change must preserve.
/loop
Repeat act, test, inspect evidence and correct—then stop at completion, the attempt limit or an approval boundary.
/review
Evaluate the result against approved requirements and report evidence-backed defects by severity.
/handoff
Record exact state, revision, next action, risk, secret boundary and rollback for the next operator.
ChatGPT
Ongoing writing, study, research and mixed-media projects
Official reference: OpenAI: Projects in ChatGPT ↗Projects
Keep related chats, files and project instructions together for ongoing work.
Project instructions
Set project-specific guidance for tone, format and working method without changing global instructions.
Source boundary
State which supplied files and approved web sources may support the answer.
Codex
Repository work that needs instructions, evidence and code review
Official reference: OpenAI: Codex developer commands ↗AGENTS.md
Store repository and directory-scoped working agreements in version control.
/compact
Replace an overlong transcript with a concise continuation summary when context grows.
/review
Start a review of uncommitted changes or compare the current work against a base branch.
Claude Code
Long-running coding work with project rules and bounded permissions
Official reference: Anthropic: Claude Code memory ↗CLAUDE.md
Provide concise project standards, common commands and architecture context.
/init
Generate or improve a starting project instruction file after repository analysis.
Plan before mutation
Inspect first, name affected files and checks, then edit only the approved scope.
Gemini + Gemini CLI
Multimodal analysis, long context and terminal-based project work
Official reference: Google: Gemini CLI commands ↗GEMINI.md
Load hierarchical project instructions into Gemini CLI sessions.
/memory show
Inspect the concatenated context currently supplied by instruction files.
@path
Attach the contents of a named file or directory to the current prompt.
GitHub Copilot
Repository-aware assistance, code review and team conventions
Official reference: GitHub: Custom instruction support ↗.github/copilot-instructions.md
Provide repository-wide instructions in the Copilot surfaces listed in GitHub’s compatibility matrix.
Path-specific instructions
Where supported, apply specialist rules only to matching files instead of overloading global context.
Acceptance-first task
Name the desired behaviour and required tests before requesting implementation.
Kimi
Long documents, reference-led synthesis and structured writing
Official reference: Moonshot AI: Prompt best practices ↗Reference-only answer
Tell Kimi to answer from supplied reference text and say when the answer is absent.
Chunk, summarise, combine
Summarise long material in bounded sections before producing a summary of summaries.
Delimited brief
Separate context, task, constraints and output format with visible delimiters.
Perplexity
Current-web research that should retain sources and project context
Official reference: Perplexity: Projects (current page on a legacy permalink) ↗Projects
Keep searches, files, instructions and collaborators inside one persistent workspace.
Prioritised domains
Add trusted links or domains when the research must privilege named sources.
Claim-source table
Return each material claim beside its source and note any unresolved disagreement.
Grok
Web and X research combined with code-backed analysis
Official reference: xAI: Tools overview ↗web_search + x_search
Combine web sources with X posts when both evidence classes are relevant.
code_interpreter
Use code execution for calculations and analysis instead of prose arithmetic.
Evidence classes
Separate official facts, reported claims and social signals in the final answer.
DeepSeek API
OpenAI-compatible API workflows with explicit output contracts
Official reference: DeepSeek: Models and capabilities ↗Thinking mode
Choose thinking or non-thinking behaviour deliberately for the task and latency budget.
JSON output
Use the documented structured-output capability when downstream code requires JSON.
Validate before routing
Test the selected model, schema and tool behaviour before changing production traffic.
NVIDIA NIM
Self-hosted or enterprise inference with observable health
Official reference: NVIDIA: NIM LLM API reference ↗GET /v1/models
Inspect the model identifier currently served before sending inference requests.
GET /v1/health/ready
Route traffic only after the loaded model reports readiness.
Compatibility contract
Record the endpoint format, model ID, tool support and fallback before deployment.
Microsoft Copilot
Work grounded in Microsoft 365 files, pages and notebooks
Official reference: Microsoft: Copilot Notebook instructions ↗Notebook instructions
Set persistent topic, tone and response-format guidance for a Copilot Notebook.
References
Ground the notebook in the relevant Word, Excel, PowerPoint, PDF or OneNote material.
Pattern source ↗Decision memo
Return recommendation, evidence, trade-offs, owner and next action in one reusable page.
Vibe
Professional chat, coding and knowledge-backed agents
Official reference: Mistral AI: Vibe overview ↗Vibe Work / Code / Chat
Choose the product mode that matches professional work, coding or turn-based chat.
Library-grounded agent
Connect an agent to an approved Library when its answers must use a controlled knowledge set.
Pattern source ↗Source-safe brief
Name allowed references, prohibited assumptions and the required evidence in the output.
A shortcut improves one move
Build a Prompt Pack →A complete Prompt Pack protects the whole outcome.
Use the reference to choose product-native controls. Use Maviter when the job needs decisions, phases, checks and a saved handoff.