How to use the RenderDoc GPU capture analyzer agent skill

What the skill lets the agent do with a RenderDoc capture

Once installed, the agent can open a RenderDoc .rdc graphics frame capture and answer questions about it without a GUI. The skill is a headless command-line harness, so the agent runs commands and reads the results rather than clicking through the RenderDoc interface.

The skill covers capture inspection (metadata, sections, thumbnails, format conversion), the action tree (listing, searching and filtering draw calls, clears, dispatches and markers), textures (listing, inspecting, exporting and pixel picking), and pipeline state at any event. It can also export shaders in human-readable form, read constant buffers, decode vertex shader inputs and outputs, read buffer data, and enumerate GPU performance counters.

• Search for passes by name, such as a shadow pass
• Export textures and render targets as PNG, JPG, BMP, TGA, HDR, EXR or DDS
• Export shaders as HLSL, GLSL or disassembly, depending on the graphics API
• Compare pipeline state between two events, in one capture or across two captures

What you need before you install the skill

The skill inspects existing captures only. You need RenderDoc installed with its Python bindings, and the renderdoc module must be on your PYTHONPATH. You also need a capture file (.rdc) to analyse. The skill does nothing useful without one.

The Python package depends on click (version 8.0 or later) and prompt-toolkit (version 3.0 or later). The install command pulls the Python package in, but the RenderDoc bindings come from your RenderDoc installation.

• RenderDoc installation that provides its Python bindings (the renderdoc module on PYTHONPATH)
• Python package cli-anything-renderdoc
• click>=8.0
• prompt-toolkit>=3.0
• A RenderDoc .rdc capture file

Install the skill for Claude

Download the skill ZIP from the page where you found this guide. The ZIP contains SKILL.md. Unzip it into one of these locations: ~/.claude/skills/ for your user account, or a project's .claude/skills/ folder if you want the skill limited to one project.

If you use the Claude apps instead, upload the skill under Settings, then Capabilities, then Skills.

The skill also needs the Python package. Install it with pip, then confirm that RenderDoc's Python bindings can be found on PYTHONPATH before you ask the agent to open a capture.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/ folder
• Claude apps: Settings, then Capabilities, then Skills, then upload

Ask the agent for analysis in plain language

You do not need to type the commands yourself. Describe what you want to learn about the capture and point the agent at the .rdc file. The agent chooses the matching commands. Good requests name the capture, the question and, where you know it, an event ID from the action list.

The commands below show what the agent will typically run. You can also run them yourself to check its work. The skill supports a --json flag on all commands, which the agent uses for machine-readable output. Use -c to name the capture file.

• 'Summarise what is in frame.rdc and how many draw calls it has.'
• 'Find the passes whose names contain Shadow and list their event IDs.'
• 'Export the fragment shader for event 42 to ./shaders/.'
• 'Compare the pipeline state between events 100 and 200 in a.rdc and b.rdc.'
• 'Save the render targets written by event 42 to ./renders/.'

Understand shader export output by graphics API

Shader export depends on the capture's API. For binary shaders the skill exports embedded source where it exists and falls back to disassembly. For text shaders it saves the raw source directly. The agent should read the file extension to know which form it has.

Direct3D 11 shaders are DXBC binaries, exported as embedded HLSL (.hlsl) or bytecode assembly (.dxbc.asm). OpenGL and GLES shaders are GLSL text (.glsl) and already readable. Vulkan shaders are SPIR-V binaries, exported as embedded GLSL (.glsl) or SPIR-V assembly (.spv.asm).

• D3D11: .hlsl or .dxbc.asm
• OpenGL / GLES: .glsl
• Vulkan: .glsl or .spv.asm

Safety notes: scripts run locally and there is no network use

The skill runs scripts on your machine. The skill's own description lists it as running scripts and not using the network. Because the agent executes commands, review the skill files and the commands it proposes before you let it run them on sensitive captures.

Captures can contain game assets and proprietary rendering code. Keep capture files and exported shaders where your team's policies allow. Diff results and preview bundles are written to disk, so check output locations before you share them.

• Scripts run on your machine; review SKILL.md and the commands the agent plans to run
• No network use is listed for this skill
• Exported shaders and textures may contain proprietary content; handle them under your team's rules

Tips for getting reliable results

Start with a summary before you dig in. Counts by type show how complex the capture is, and listing only draw calls keeps the output focused on real rendering. Pipeline state needs an event ID from the action list, so get the ID first. Pipeline diff writes a JSON file and prints only its path, so ask the agent to read that file for the details.

• Run actions summary first to understand the capture's complexity
• Use actions list --draws-only to focus on rendering work
• Get event IDs from the action list before asking for pipeline state
• Set RENDERDOC_CAPTURE to a default capture path if you analyse one file repeatedly
• Use pipeline diff with -b to compare events across two captures