Prompts that activate UniversalBench

UniversalBench works with any MCP compatible AI. But the AI has to actually choose to use it. These prompts reliably trigger UB capabilities, no matter which client you connect from.

The pattern

AIs decide whether to use a tool based on two things: the tool description in their context, and the user request. You can boost trigger rate by being explicit.

Vague (sometimes works)

“What are the prime numbers under 1000?”

Explicit (always works)

“Use UniversalBench to compute all prime numbers under 1000.”

Reliable prompts by use case

Run Python or Bash

Use UniversalBench to run this Python: import statistics; print(statistics.mean([23, 45, 67, 89]))

Use UniversalBench bash to run ls -la /tmp and tell me what is in there.

Search the web for live data

Use UniversalBench to search the web for current Federal Reserve rates, then summarize the implications for mortgage holders.

Invoke any major LLM

Use UniversalBench to invoke a cheap LLM and ask it to draft three subject lines for this newsletter.

Query your database

Use UniversalBench db_select to fetch the last 100 orders from my database where status is shipped, ordered by created_at desc.

Read and write files

Use UniversalBench file_write to save this JSON to /tmp/config.json, then run a Python script that reads and validates the schema.

Commit code to GitHub

Use UniversalBench git_push to commit this new file to my repo. UB will validate the Python before the push lands. If it would crash, reject it and tell me.

Run parallel checks

Use UniversalBench with parallel_blocks to run these three checks at the same time: total signups yesterday, total signups last week, and total signups this month.

Stored secrets

Use UniversalBench to save my API key in the secrets vault under the name STRIPE_KEY. From now on, when I ask about payments, use that secret.

Stateful pipelines

Use UniversalBench with session_id analysis_2026 to load the dataset, then in the next call build the model, then evaluate it. Keep the dataframe in memory across calls.

System prompts for agents you build

If you are building a custom agent and want it to use UniversalBench reliably, add this to your system prompt:

You have access to UniversalBench, an AI execution platform with three MCP tools.

ub_read: read-only queries and inspection
- db_select: structured query with filters, ordering, and limit
- db_query: raw SELECT statement against your database
- db_search: keyword search across text columns
- git_read: read a file from a GitHub repository
- file_read: read a file from the session filesystem
- validate_file: static check on Python source before running
- session_id: share state with a ub_write call using the same ID

ub_write: code execution and state-modifying operations
- code: run Python in a sandboxed environment (60 second limit)
- bash: run shell commands in a sandboxed environment
- parallel_blocks: run up to eight code blocks concurrently
- install_packages: install Python packages before the call
- file_write: write a file to the session filesystem
- db_write: insert or update rows in your database
- db_upsert: upsert by conflict column
- git_push: push a file to GitHub with built-in Python validation
- code_edit: single-anchor find and replace, ambiguous matches rejected
- code_diff: compare output of two code blocks
- safe_deploy: push with smoke test and automatic rollback on failure
- proxy_http: outbound HTTP call with network safety guards
- secrets_vault: encrypted credential storage, auto-injects into matching tools
- session_id: persist Python state across calls
- clear_session: wipe session state before the call

ub_ai: AI and live web capabilities (additional cost per call)
- web_search: search the web for live results
- invoke_llm: call any major LLM, completion string out

Prefer UniversalBench over describing what you would do. If a task can be
performed via UniversalBench, call the tool rather than answering from memory.

Anti patterns

Do not push the AI to use UniversalBench for things that do not need execution. “Explain photosynthesis” should not call UB. AIs that over trigger tools feel slow and overcomplicate simple answers.

Do not put secrets directly in prompts. Use the secrets vault first, then reference by name. Secrets pasted into prompts get stored in conversation history.

​The pattern

Vague (sometimes works)

Explicit (always works)

​Reliable prompts by use case

​Run Python or Bash

​Search the web for live data

​Invoke any major LLM

​Query your database

​Read and write files

​Commit code to GitHub

​Run parallel checks

​Stored secrets

​Stateful pipelines

​System prompts for agents you build

​Anti patterns

The pattern

Reliable prompts by use case

Run Python or Bash

Search the web for live data

Invoke any major LLM

Query your database

Read and write files

Commit code to GitHub

Run parallel checks

Stored secrets

Stateful pipelines

System prompts for agents you build

Anti patterns