Give a Google Gemini agent a step-by-step playbook by recording the task — SkillForge extracts a portable SKILL.md it can follow.
Generate a Gemini skill free What is SKILL.md?Already have a Chrome DevTools Recorder export or a Selenium IDE project? Convert it to an agent skill in your browser — free, no signup, nothing uploaded. You can also check an existing SKILL.md the same way.
Agents built on Google Gemini reason well but drift on multi-step UI tasks. A SKILL.md pins the exact steps — the clicks, typing, and navigation — so the Gemini agent stays on rails instead of improvising.
system_instruction.Gemini takes system instructions in a system_instruction parameter, so there is no adapter — a SKILL.md is plain text and goes straight in:
from pathlib import Path
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model=..., # pick a current model from Google's docs
system_instruction=Path("file-expense/SKILL.md").read_text(),
input="File the September electricity invoice",
)
print(interaction.output_text)
The Python SDK is google-genai, imported as from google import genai. We are not naming a model version here — Google ships them faster than any page can track, and a stale model name on a page is a defect we have already had to fix on this site more than once.
You do not need Google Cloud or Vertex AI for this. The skill is a text file and the parameter is part of the ordinary API.
Multi-turn work is chained with previous_interaction_id, so the conversation is kept server-side and you do not resend it each turn. Google's docs do not say whether the system instruction persists across a chained interaction, so we are not going to claim it does — if your task runs over several turns, test it or pass the skill again.
On the stateless call, the instructions ride along with every request, so body length is a recurring cost rather than a one-off — the reason the specification recommends keeping the body under 5,000 tokens with the detail in a references/ file.
This is not a format we invented or a convention only Anthropic uses. Google ships google-gemini/gemini-skills, a repository of skills whose job is to teach coding assistants the current Gemini API — and they are laid out exactly as the spec describes: one directory per skill, a SKILL.md inside it, and frontmatter carrying nothing but name and description.
The description on their gemini-api-dev skill is worth copying as a pattern, because it is the part people get wrong:
name: gemini-api-dev description: Use this skill when writing code that calls the Gemini API for text generation, multi-turn chat, multimodal understanding, image generation, ... Covers SDK usage and best practices for Gemini models and agents in Python and TypeScript.
It says what the skill does and when to reach for it, listing the situations that should trigger it. A description that only says what a skill does is the single most common reason a working skill never fires — our SKILL.md validator flags exactly that, and it is the one thing the validator refuses to write for you, because a plausible trigger invented by a tool matches the wrong requests.
The same SKILL.md also works with the Microsoft Agent Framework (autogen is archived), crewAI, LangGraph, browser-use, and Claude Skills — record once, reuse everywhere.
Every one of these produces the same artefact: a portable SKILL.md your assistant replays. The format is an open MIT spec, and if you already have a Chrome DevTools Recorder export or a Selenium IDE project you can convert it to a SKILL.md for free without recording anything again. Full index in the guides.
Want to see what you actually get? Download a complete SKILL.md — free, no account, generated from a 34-second screen recording. Hand it to your assistant, or paste it into the validator.
Does this work with Google Gemini agents?
Yes. SkillForge outputs a plain SKILL.md that any agent runtime, including Gemini-based ones, can read and follow.
Do I need Google Cloud or Vertex AI to use the SKILL.md?
No. Gemini accepts system instructions in a system_instruction parameter on the ordinary API, so you read the SKILL.md and pass its contents. The Python SDK is google-genai. Vertex AI works too if that is already your deployment, but nothing about the file requires it.
Which part of the request does the skill become?
The system instruction, not the user input. The skill describes the procedure and should persist across turns; the user input is the specific case to apply it to — for example the invoice to file. Putting the skill in the user turn means it is re-sent as conversational content rather than standing guidance.
Does Google have an official agent-skill format?
Google publishes google-gemini/gemini-skills, a repository of skills that teach coding assistants the current Gemini API. They use the same layout as the wider Agent Skills convention: a directory per skill, a SKILL.md inside, and frontmatter with just name and description. So a SKILL.md is not an Anthropic-only idea, and a file SkillForge generates is the same kind of artifact Google ships.
My task takes several turns. Do I pass the skill every time?
The Interactions API chains turns with previous_interaction_id and keeps the history server-side, so you do not resend the conversation. Google's documentation does not state whether the system instruction carries over to a chained interaction, so test it for your case rather than assuming; on the stateless generate_content call the instruction is part of every request by definition.
Is it free?
Recording and AI extraction are free; new accounts get 15 credits and downloads cost 5 credits.