Summary

Adds two new SKILL.md files to skills/ that teach AI coding agents (Claude Code, Cursor, Codex, Gemini CLI, etc.) how to use AgenticFlow's built-in features and model selection:

• agenticflow-built-in-credits — Credits-first approach: use agenticflow_generate_image, web_search, web_retrieval, and built-in blueprints before considering BYOK (Bring Your Own Key). Includes the decision framework (try built-in → iterate prompts → evaluate → extend only if needed), credit cost tables, blueprint-first agent setup, and explicit handoff to agenticflow-mcp for BYOK when the user is genuinely unsatisfied.

• agenticflow-llm-models — Model selection guide with agenticflow/glm-4.7-flash as the default (best all-rounder with reasoning + tools at 18s), agenticflow/gemini-2.5-flash-lite as the fast alternative (13s, no reasoning overhead), plus the top 5 recommended models, models to avoid (Claude 3.5 Haiku for logic, deprecated Gemini 2.0 Flash, flaky GPT-OSS variants), reasoning token configuration via af schema agent --field model_user_config, and the finish_reason=length fix (increase max_tokens to ≥4000 for GPT-5 Nano and GLM-4.5 Air).

Why these skills

Without skills, AI agents responding to AgenticFlow questions:

• Invent non-existent CLI commands (af platform-catalog models, af credits balance, af agent connect-tool, --reasoning-effort)
• Recommend wrong/deprecated models (gpt-4o-mini over glm-4.7-flash; gemini-2.0-flash which is deprecated)
• Skip prompt engineering and jump straight to BYOK when the user mentions image generation
• Can't troubleshoot AgenticFlow-specific errors like finish_reason=length with model-specific max_tokens guidance

Model default: glm-4.7-flash first, gemini-2.5-flash-lite second

The default model is agenticflow/glm-4.7-flash — it has built-in reasoning (931 tokens), tool use, and at 18s is only 5s slower than Gemini 2.5 Flash Lite while delivering much stronger results for agents that need to think and call tools. Gemini 2.5 Flash Lite (13s) is recommended as the fast alternative when you don't need reasoning.

Evaluation results

Both skills were tested using the OpenCode Skill Creator — the open-source tool for creating, evaluating, and iteratively improving AI skills with automated benchmarks. 3 test prompts per skill, run with and without the skill:

The built-in-credits skill went through 3 iterations (85% → 93% → 100%) with improvements driven by eval failures:

• Iteration 1→2: Added explicit agenticflow-mcp skill reference for BYOK extension steps
• Iteration 2→3: Strengthened "blueprint first" instruction — agent now uses af agent init --blueprint content-creator instead of creating custom agents from scratch

Files added

Both skills follow the existing convention (SKILL.md only, matching agenticflow-agent, agenticflow-workforce, agenticflow-mcp).

Related

• Follows the pattern established by the existing 3 skills in skills/
• Supports the AGENTS.md goal of keeping skills next to the commands they document
• Built with OpenCode Skill Creator — open-source tool for creating, evaluating, and iteratively improving AI skills using automated benchmarks