agentsclimarketplace

Agent optimization

Skill abhishekgahlot2/agent-optimization/skills/agent-optimization

Prompt, checklist, and Agent Skill for optimizing LLM agents without benchmark hacks.

Install
npx -y skills add abhishekgahlot2/agent-optimization --skill agent-optimization

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when optimizing, reviewing, or debugging LLM agents for latency, cost, reliability, delivery quality, benchmark integrity, prompt bloat, retry loops, tool-call waste, fast paths, or eval regressions. Apply the Agent Optimization Prompt: measure honestly, use real traces, exhaust deterministic levers before model/prompt tweaks, and verify on the real serving path.

SKILL.md

2.3 KB, as published. Nobody here has run it

Agent Optimization

Use this skill when the user wants to improve an LLM agent without reward-hacking the metric.

Core Rule

Never optimize the number. Optimize the thing the number measures.

If a metric improves and user-visible delivery does not, say so.

Workflow

  1. Establish the baseline:

    • metric
    • harness
    • sample size
    • trace source
    • model/provider/version if relevant
    • delivery and correctness gates

  2. Inspect real traces or failures before proposing fixes.

  3. Decompose the budget:

    • model generation and deliberation
    • turns and round trips
    • tool/database/service time
    • fixed overhead
    • retries and failed attempts

  4. Rank deterministic levers before prompt/model changes:

    • caching
    • merged round trips
    • parallel tool calls
    • schema/semantic layers
    • typed intents
    • compilers/generators
    • validators
    • server-side budgets

  5. Check unsupported asks:

    • If the action space cannot express the answer, require fallback, partial output with a note, or honest refusal.
    • Never accept silence as a valid degradation mode.

  6. Verify on the real serving path.

  7. Report:

    • measured before/after
    • method
    • gates passed
    • residual classes
    • behavior changes
    • rollout risk

Reference Prompt

When a full copy-paste prompt is useful, read:

  • references/agent-optimization-prompt.md

When a compact review gate is enough, read:

  • references/agent-optimization-checklist.md

Output Shape

Prefer this structure:

## Verdict

PASS / FAIL / NEEDS DATA

## Evidence

- Baseline:
- Candidate:
- Harness:
- Data:
- Gates:

## Bottleneck

What actually dominates the budget.

## Fix Priority

1. Deterministic fix first
2. Next deterministic fix
3. Prompt/model change only if model-bound

## Verification Plan

How to prove the change without weakening gates.

## Risks

What may regress, what was not tested, and what telemetry must watch.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.