AI Agent Workflow Collection

Chat with gpt-oss-120b, ZeroGPU's largest open-weight model (117B MoE, 131K context). Use when a task needs more capability than the 1.2B edge models can give — longer documents, multi-step instructions, harder general-knowledge questions — but still does not warrant Claude.

Fastest, cheapest chat reply via Liquid AI's edge model (LFM2.5-1.2B-Instruct). Use when the user wants a quick, single-turn answer that does not need Claude-level reasoning, prior conversation context, or code generation, and explicitly favors speed or cost over quality. Optional system instructions via -i.

Chat with qwen3-30b-a3b-fp8, a 30.5B MoE model with strong multilingual coverage (100+ languages, 32K context). Use for non-English prompts, translation-adjacent tasks, or mid-weight questions the 1.2B edge models handle poorly.

Chat with ZeroGPU's Thinking variant (LFM2.5-1.2B-Thinking), which returns visible reasoning alongside the answer. Use when the user wants the model's reasoning shown, or asks a short logic/math/word-problem question that benefits from step-by-step output.

Classify a domain name against the IAB taxonomy without fetching the page (zlm-v1-iab-domain-classifier). Use when the user has a hostname or URL rather than article text — bidstream enrichment, allow/deny-list scoring, "what is example.com about?".

Classify text against the IAB content/audience taxonomy. Use when the user asks for IAB categories, ad-taxonomy labels, or "what topic is this article about" against a standard taxonomy.

Enriched IAB classification — returns audience categories plus topics, keywords, and inferred intent. Use when the user wants richer ad/audience signals than plain IAB labels (e.g. "give me topics, keywords, and intent for this passage").

Multi-axis classification using a JSON schema mapping categories to allowed labels (gliner2-base-v1). Use when the user wants to classify text along several dimensions at once, e.g. "classify by sentiment and topic" with explicit label sets per axis.

Zero-shot classification against a caller-supplied list of candidate labels (deberta-v3-small). Use when the user wants to classify text into a custom set of labels they provide (e.g. "is this positive, negative, or neutral?", "tag this as bug, feature, or question").

Custom-label named-entity recognition (gliner2-base-v1). Use when the user wants to extract entities with their own labels — people, organizations, locations, products, dates, or any caller-defined entity types — from a passage.

Schema-driven structured JSON extraction (gliner2-base-v1). Use when the user wants to pull specific named fields out of free text into a JSON object — contact info, invoice details, order data, profile attributes — defined by a per-field type/description schema.

Extract PII entities from text (gliner-multi-pii-v1). Use when the user wants to find personally identifiable information — names, emails, phones, addresses, financial identifiers — grouped by category, without modifying the source text.

Generate natural follow-up questions a reader would ask next about a passage (zlm-v1-followup-questions-edge). Use when the user wants suggested next questions, "people also ask" style prompts, or conversation continuations for an article, answer, or chat turn.

Detect and mask PII in-line in the text, replacing it with label placeholders like [PERSON] and [EMAIL]. Use when the user asks to redact, scrub, mask, anonymize, or sanitize a passage before sharing or logging it.

Sign in to ZeroGPU and persist the API key + Project ID. Use when the user asks to log in, authenticate, or set up ZeroGPU credentials for the first time.

Route cheap AI tasks (classification, summarization, entity/JSON extraction, PII redaction/extraction, follow-up question generation, small-model chat) to ZeroGPU small/nano models instead of spending host-model tokens. Invoke whenever the user asks for one of these tasks and the input is plain text.

Summarize a plain-text passage you already have in hand — an email, transcript, meeting notes, article body, chat log — using ZeroGPU's llama-3.1-8b-instruct-fast edge model. Use when the user asks to summarize, condense, TL;DR, or give the gist of text that is already in the conversation or pasted into the prompt. Not for URLs, YouTube links, podcasts, PDFs, or local files — the bundled summarize skill handles those.