Die Oberfläche ist lokalisiert; Skill-Beschreibungen und -Dokumente werden per KI automatisch in Ihre Sprache übersetzt. Code, Befehle, Pfade und Fachbegriffe bleiben im Original, maßgeblich ist das Dokument des Autors.
Skills entdecken
Entwicklung · SKILL.md

techsmith-performance-tuning

techsmith-performance-tuning

Optimize Snagit capture and Camtasia project/export throughput from measured workstation bottlenecks and safe concurrency limits. Use when queues are slow, unstable, or resource-bound. Trigger with "TechSmith performance", "Camtasia export tuning", or "Snagit capture throughput".

Independent GitHub snapshot · Automated classification · Original documentation governs

Skill-Dokumentation

GitHub

name: techsmith-performance-tuning description: >- Optimize Snagit capture and Camtasia project/export throughput from measured workstation bottlenecks and safe concurrency limits. Use when queues are slow, unstable, or resource-bound. Trigger with "TechSmith performance", "Camtasia export tuning", or "Snagit capture throughput". argument-hint: "[workload-manifest] [workstation-alias]" allowed-tools: Read, Glob, Grep, WebFetch, Write, Edit version: 1.6.0 author: Jeremy Longshore jeremy@intentsolutions.io license: MIT tags:

  • desktop-automation
  • techsmith
  • performance model: inherit effort: medium compatibility: Designed for Claude Code; TechSmith product execution requires an approved Windows or macOS desktop as documented

TechSmith Workstation Performance Tuning

Overview

This skill improves throughput without assuming server-style horizontal scale. It measures one representative workload, separates capture from export bottlenecks, keeps active Camtasia data local, and changes one resource or concurrency control at a time.

Prerequisites

  • A representative, non-sensitive capture or project workload and acceptance criteria
  • Baseline wall time, CPU, GPU, memory, disk, output size, and failure rate
  • Product/version, hardware, local scratch path, and export preset
  • A rollback point and permission for any preference or driver change

Tool Discipline

Use Read, Glob, and Grep to inspect local scripts, manifests, logs, and tests. Use WebFetch only for current primary TechSmith documentation. Use Write or Edit only after confirming the target repository file and approval boundary.

Current Contract

  • Serialize interactive Snagit captures per desktop unless validated isolation proves otherwise.
  • Keep Camtasia project media, scratch, and export output on a fast local non-synced drive while active.
  • Bound parallel exports by measured CPU/GPU/memory/disk headroom and licensed workstation capacity.
  • Do not trade correctness, media validation, or workstation stability for raw queue throughput.

Licensing and Authentication

TechSmith desktop activation is not API authentication. Resolve individual sign-in versus business-key or approved offline activation before execution. Redact all keys, account identifiers, activation artifacts, and sensitive endpoint details.

Instructions

  1. Capture a cold and warm baseline for one fixed workload with product/version and preset pinned.
  2. Identify the dominant constraint: interactive desktop, source media decode, CPU, GPU, memory, local disk, codec, or operator wait.
  3. Set a conservative worker limit and queue backpressure from measured peak headroom.
  4. Apply one reversible change such as local scratch, proxy behavior, preset correction, or worker count.
  5. Repeat the identical workload and compare duration, utilization, output properties, and failure rate.
  6. Keep only changes that improve the declared service objective without a quality or stability regression.

Approval Boundaries

Do not edit projects on cloud/network storage, disable security software, force undocumented command flags, or launch duplicate exporters to manufacture concurrency.

Output

Return baseline and candidate metrics, bottleneck, change, before/after delta, quality checks, safe worker limit, rollback, and monitoring threshold.

Error Handling

ConditionResponse
Workload is not reproducibleFreeze inputs and preset before drawing conclusions.
Disk saturationReduce workers and stage active files on approved local storage.
Output differsReject the optimization and restore the baseline configuration.
Crash rate increasesRollback immediately and preserve redacted diagnostics.

Examples

The example below shows the minimum redacted evidence expected from a successful invocation of this operator workflow.

workload=training-20m; workers=1->2; duration=-28%; peak_mem=82%; quality=pass; decision=keep

Resources

Gespeicherte Skills