Write a prompt in BRIEF, then prove it works.

“I wrote a prompt that seems to work. How do I find out if it really does?”

Turns a plain-language ask into the five BRIEF blocks, then evaluates it: pass rate, a bar per criterion, a consistency score, and the delta against the naive version.

RUN THE OPERATOR SYSTEMADVANCED✓ APPROVED 4/4
EXAMPLE PROMPTS
>Turn this ask into a BRIEF prompt for scoring inbound leads
>Eval this prompt and give me a pass rate for each criterion
>How much better is my BRIEF prompt than the naive version? Run the comparison
ABOUT THIS SKILL

What this skill does

Two jobs: brief a system with the BRIEF framework, and prove it works with the eval harness. BRIEF is the system. SCRIPT is each seat inside it. Most agent setups fail in production because nobody measured them, so the second half is the point.

How it works

  1. 1
    1. **Brief request** ("brief this agent", "set up the loop") means run BRIEF, deliver

    the five blocks plus one SCRIPT per role, clean. A single prompt with one seat is a SCRIPT, not a BRIEF: hand it to prompt-architect. 2. Eval request ("does this work", "what's the success rate") means write a config from the Finish line, run brief_eval.py, deliver the scorecard. 3. Both means write the BRIEF, then offer to eval it. Anything about to run across many rows, deals, or accounts without a person watching should be evaluated before it ships.

Output format

  • Pass rate = average fraction of checks passed across all runs.
  • Consistency = 1 minus the spread across datasets. Low consistency means the prompt is unreliable even if its average looks fine. Tighten the Finish line.
  • Per-criterion bars show which requirement leaks. A low hard_stop bar means the model keeps adding summaries; a low claims_supported bar means it's fabricating. Fix the specific block, rerun, watch the bar move.

This is the same loop as a code test suite: define success (Finish line), run the suite (eval), fix the weakest part, rerun.

RELATED SKILLS
NO SETUPADVANCED · Standalone

agent-contract

Turn a named agent job into the contract that decides whether it can be evaluated at all: the input line, the output schema, the must-always rules, the must-never list, and the abstain state.

RUNS ON WHAT YOU PASTEAPPROVED · INSTALL →
NO SETUPADVANCED · Standalone

agent-seam

Define the handoff between two agents so a chain holds up in production: the payload schema both sides agree on, who validates it, how a refusal travels instead of turning into a null, what keys a record so a batch run cannot cross accounts, and where the checkpoint sits before anything expensive or irreversible.

RUNS ON WHAT YOU PASTEAPPROVED · INSTALL →
ALL RUN THE OPERATOR SYSTEM SKILLS →