Agent skill
Singularity Compute publishes a Super Decider skill for coding agents. With it installed, an agent such as Claude Code, Codex, Cursor or Gemini CLI knows when a feature needs a decision instead of generated text, how to write the questions, how to read the probabilities in code and how to call the API.
The skill works offline. It bundles:
- An API reference and a question-design guide.
- Patterns such as routing, confidence gating, guardrails, reranking, extraction and agent tool gating.
- Ten runnable Python examples.
- A guide to moving an existing Jev integration to Super Decider.
- Two scripts:
check_request.pyto validate a request body, anddecide.pyto send one.
Source: Singularity-Compute/skills on GitHub, Apache-2.0.
Install​
Claude Code​
Install it as a plugin:
claude plugin marketplace add Singularity-Compute/skills
claude plugin install super-decider@singularity-compute
Codex, Cursor, Gemini CLI and other agents​
Install it with skills.sh and pick your agent when prompted:
npx skills add Singularity-Compute/skills --skill super-decider
Add -g to install it for all projects.
Manually​
Copy the skills/super-decider/ folder into your agent's skills folder, for example ~/.claude/skills/super-decider/.
Point it at ASI:Cloud​
The skill's examples and scripts read the API key from API_KEY, and the server address from SUPER_DECIDER_URL. By default they call Singularity Compute's own endpoint. To use ASI:Cloud, set both:
export API_KEY="your-asi-cloud-api-key"
export SUPER_DECIDER_URL="https://inference.asicloud.cudos.org"
Create the key on the API keys page. The scripts add /v1/systemone to the address themselves.
When you ask the agent to write code, tell it to use the ASI:Cloud endpoint, https://inference.asicloud.cudos.org/v1, so the code it writes doesn't use the skill's default address.
Use it​
Ask your agent for a feature in plain language. For example:
Use Super Decider to route incoming support tickets by team, with human review when it is unsure. Call it at https://inference.asicloud.cudos.org/v1.
Add a guardrail that checks every message to and from our chatbot for prompt injection and personal data.
We call Jev to classify leads. Switch the integration to Super Decider.
The agent loads the skill when the task matches. In Claude Code you can also call it directly with /super-decider:super-decider.
Examples in the skill​
| Example | Pattern |
|---|---|
route_tickets.py | Routing, with speculative follow-up questions and a confidence gate to human review |
guardrails.py | Screening LLM input and output: one yes/no per hazard, then block, review or pass in code |
rerank.py | Reranking retrieved passages with one score each, plus an "is it answerable" check |
extract_values.py | Copying values out of text with span questions, then validating them in code |
verify_claims.py | Checking claims or citations against a source, with thinking on |
composite_score.py | Scoring a lead on several dimensions and weighting them in code |
agent_tool_gate.py | Deciding whether an agent's tool call runs, asks the user first, or is refused |
model_router.py | Sending each request to the cheapest LLM that can handle it |
multilabel_moderation.py | Labels that can apply together: one yes/no per label |
thinking_cascade.py | A fast pass for everything, then a second pass with thinking only for unsure items |
Run one from the skill folder with python3 examples/<name>.py, after setting the environment variables above.
Related​
- Migrate from Jev: the changes the skill makes when it moves a Jev integration.
- Patterns: the same patterns, explained in these docs.