skills.rest
AnderHonorato
AnderHonoratoCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

builtin-llm-models

Call Manus built-in LLM proxy models with correct request shapes and batch scripts.

Updated Aug 28, 2026
Search Tags:#api#llm#reasoning#batch processing#structured output#openai-compatible#model catalog
One-click install
npx skills add https://github.com/AnderHonorato/Mem-rias-IA---Infinity --skill builtin-llm-models-anderhonorato
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: builtin-llm-models
Source: https://github.com/AnderHonorato/Mem-rias-IA---Infinity/tree/main/Manus/Skills/builtin-llm-models
Command: npx skills add https://github.com/AnderHonorato/Mem-rias-IA---Infinity --skill builtin-llm-models-anderhonorato

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires openai, and includes scripts (resource) components.

What problem does it solve? Calling the Manus built-in LLM proxy incorrectly wastes tokens and causes silent failures: each model family (GPT, Claude, Gemini) requires different thinking/reasoning parameters and different max-token fields, and the model catalog changes over time. This Skill provides the verified request shapes, live catalog discovery steps, and a concurrent batch driver so bulk LLM jobs run correctly the first time. ## Core Features & Use Cases - Live model catalog discovery: Fetch current model IDs, pricing, and capabilities from the /models endpoint instead of relying on stale hardcoded lists. - Family-specific request shapes: Correct thinking/reasoning parameters and max-token handling for GPT (reasoning.effort, max_completion_tokens), Claude (thinking budget_tokens with max_tokens), and Gemini (max_tokens only). - Structured output, vision, and tool calling: JSON-schema strict output, image inputs, and function calling via the OpenAI-compatible API in both sandbox Python and webdev TypeScript (invokeLLM). - Batch processing script: scripts/llm_batch.py runs concurrent, retrying bulk jobs for summarization, translation, classification, and extraction with optional JSON-schema output. - Use Case: Translate 5,000 product descriptions by running llm_batch.py with gpt-5-mini, then re-run only schema-validation failures on claude-sonnet-4-6 to cut costs. ## Quick Start Ask the agent to fetch the live model catalog and run a batch summarization of your input file using the llm_batch.py script with gpt-5-mini.

Frequently Asked Questions about builtin-llm-models

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I call the Manus built-in LLM from a Python script?▼

Use the openai Python SDK with OpenAI(), which auto-reads the preconfigured OPENAI_API_KEY and OPENAI_API_BASE environment variables. Call client.chat.completions.create with a model ID from the live /models catalog.

How do I run batch LLM jobs like bulk summarization or translation?▼

Use the bundled scripts/llm_batch.py driver, which reads inputs from a file, sends each to a chosen model with a prompt template, and writes JSONL results. It supports concurrency, retries, and optional strict JSON-schema output.

Which built-in model should I use for cheap bulk tasks?▼

Use gpt-5-mini as the default cheap workhorse for translation, summarization, and classification at scale, dropping to gpt-5-nano for very large volumes. Reserve claude-opus-4-7 or gpt-5.5 for judging and hardest reasoning tasks.

Why does Gemini return content null with finish_reason length?▼

This happens when max_completion_tokens is passed to Gemini models on the proxy. Switch to max_tokens (for example 16384) or omit the parameter entirely, since Gemini reasons by default.

Why does Claude return a 400 Bad Request when thinking is enabled?▼

Claude requires max_tokens to be strictly greater than the thinking budget_tokens value. For example, with budget_tokens of 2048, set max_tokens to at least 2049.

Can I use the built-in LLM in a webdev project?▼

Yes, import invokeLLM and listLLMModels from server/_core/llm in server-side TypeScript code. Credentials are injected via BUILT_IN_FORGE_API_URL and BUILT_IN_FORGE_API_KEY, and calls deduct project credits.

Related Skills

Explore More Software Engineering Skills→
Simplified-ReasoningSimplified-Reasoning
58

json-translator

Translate specified text fields in JSON files via external APIs.

Official
Intermediate
PatientVibesPatientVibes

human-review-loop

Route low-confidence agent outputs to asynchronous human review queues.

Community
Advanced
JansenAnalyticsJansenAnalytics
5

model-usage

Summarize per-model usage and costs from CodexBar local logs.

Community
Intermediate
duany049duany049

/init

A task and/or a project management system for organizing and tracking your work, optionally integrated with your calendar and email via plugins, connectors, and APIs.

Community
Advanced