- Home
- Skills
- Agents & Orchestration
- rsi skill
rsi skill
Automated evaluation and optimization loop that evolves agent skills through recursive testing and blinded grading.
$5
/moSubscription · cancel anytime
Secure checkout via Stripe
- Always the latest version
- Delivered live to your agent via MCP
- Cancel anytime, access ends at period end
Works with the AI Agents you already use
rsi skill
Example session with this skill installed
rsi-skill run against ./weather-agent until no major accuracy errors max 3 iterations $10 limit
- Read your context and instructions
- Compiled the rsi
- Generated the document
I have completed the recursive self-improvement loop for the weather-agent skill. The process ran for 3 iterations, evaluating the skill against 15 unique test cases.
RSI Execution Summary
| Metric | Value |
|---|---|
| Final Champion | v2 (Promoted in Iteration 2) |
| Goal Status | Met (No major accuracy errors found in last 5 cases) |
| Total Iterations | 3 |
| Total Spend | $6.42 / $10.00 |
| Findings Fixed | 4 (Incorrect unit conversion, missing city fallback, API timeout handling) |
| Remaining Issues | 1 (External: API rate limiting on free tier) |
Performance Trajectory
The baseline (v0) failed 40% of cases due to temperature unit confusion. v1 introduced a unit-normalization script but regressed on location parsing. v2 (the current champion) corrected the parsing logic and successfully handled all 5 fresh cases in the final evaluation iteration.
Changes Applied
The following improvements were written to weather-agent/SKILL.md and weather-agent/utils.py:
- Added explicit instruction to check for Celsius/Fahrenheit markers in the user prompt.
- Implemented a regex-based coordinate extractor to handle ambiguous city names.
- Added a retry loop with exponential backoff for
weather-apicalls.
Next steps
- Review the uncommitted changes in
./weather-agentusinggit diff. - Run
rsi-skill reportto view the full HTML evaluation report with transcripts. - Commit the changes if the benchmarks in
REPORT.mdmeet your requirements.
rsi-skill.pdf
PDF · document
Example file from a real run - the skill writes it into your workspace.
Subscription · always the latest version
- Secure checkout via Stripe
- Cancel anytime
- Delivered live via MCP
About this skill
Recursive self-improvement is hard to manage manually. Testing an agent skill across diverse scenarios, grading its outputs objectively, and iterating on the prompt or logic often leads to overfit solutions or ballooning costs.
What it does
- Autonomous Iteration executes a continuous loop of profiling, case generation, headless execution, and evidence-based grading.
- Champion-Challenger Gating ensures only changes that outperform the current best version on fresh, held-out test cases are promoted.
- Dynamic Case Generation creates new evaluation scenarios for every iteration to prevent the agent from simply "teaching to the test."
- Sandbox Execution runs skill candidates in isolated environments to safely benchmark performance and side effects.
- Resource Management tracks token usage, time, and dollar spend against user-defined limits to prevent runaway loops.
How it works
- Initialize the workspace and profile the target skill to establish a baseline and measurable success goal.
- Generate a batch of fresh, unseen test cases based on the skill's specific failure modes and objectives.
- Execute and Grade the current champion and the new challenger in parallel, using independent subagents to provide blinded, evidence-backed scores.
- Promote or Veto the challenger based on deterministic gates and aggregate performance improvements.
- Apply the final optimized version as uncommitted edits to the target project for developer review.
Frameworks & tools
The skill utilizes Python scripts for state management and budget tracking, along with general-purpose agents for profiling and grading. It interacts with the Claude CLI for headless execution and supports any agent skill structure including standalone scripts, plugins, or project-integrated skills.
Why this beats prompting it yourself
Manual iteration is prone to bias and ignores regression. This tool enforces a rigorous scientific method where every improvement must prove its worth against held-out data that the improver agent has never seen.
Use cases
- Hardening a production skill against edge cases before deployment.
- Benchmarking a skill's performance across different LLM models.
- Automating the fix-test-deploy loop for complex multi-step agent workflows.
- Identifying and documenting external API limitations that hinder skill performance.
Known limitations
Requires a sandbox-capable environment and specific CLI access. Grading accuracy depends on the quality of the judge model configured. Pre-existing environment errors in the target may be skipped.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Requires Claude CLI for headless runs. Performance depends on judge model quality. Cannot fix external API or infrastructure failures. Environment setup must allow sandbox execution.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 11 days ago
- Software
Git