### Install PocketHarness via script

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

Clone the repository and run the installation script to build, test, and install the binary to ~/.local/bin/pocket.

```bash
git clone https://github.com/yunusemrejr/pocketharness
cd pocketharness
./install.sh        # builds, tests, installs to ~/.local/bin/pocket
```

--------------------------------

### Example Configuration JSON

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

A comprehensive example of a configuration file defining providers, models, roles, hooks, and resource limits.

```json
{
  "default_model": "glm",
  "thinking": "auto",
  "providers": {
    "orcarouter": {
      "protocol": "openai",
      "base_url": "https://api.orcarouter.ai/v1",
      "key_env": "ORCAROUTER_API_KEY"
    }
  },
  "models": {
    "glm": { "provider": "orcarouter", "model": "z-ai/glm-5.3-flash" }
  },
  "roles": { "fast": "deepseek:deepseek-flash", "review": "deepseek:deepseek-flash,openrouter:z-ai/glm-5.3-flash" },
  "hooks": {
    "post_edit": ["case {file} in *.py) python3 -m py_compile {file};; esac"],
    "pre_bash": [],
    "post_bash": [],
    "stop": ["make -s test"]
  },
  "review": true,
  "autonomy": true,
  "jev": true,
  "local_lm": {
    "server": "~/llama.cpp/llama-server", "model": "~/models/Qwen3.5-0.8B-Q4_0.gguf",
    "key_file": "", "port": 18735, "threads": 4, "ctx": 4096
  },
  "tool_network": true,
  "bash_timeout": 120,
  "output_limit": 262144,
  "disk_budget_gb": 40,
  "disk_reserve_gb": 10,
  "max_file_gb": 32,
  "max_rounds": 100,
  "working_context_tokens": 96000,
  "allow_read": [],
  "allow_write": [],
  "expose_env": []
}
```

--------------------------------

### ffmpeg SDR H.264/AAC encode starting point

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/terminal-video-editing/references/encoding-verification.md

Example command for already-SDR video with a desired first audio track, if libx264 exists. Uses stream copy for video and audio, sets yuv420p, AAC 192k, and faststart. Excludes extra tracks and subtitles intentionally.

```text
ffmpeg -n -i INPUT -map 0:v:0 -map 0:a:0? -c:v libx264 -preset medium -crf 20 -pix_fmt yuv420p -c:a aac -b:a 192k -movflags +faststart OUTPUT.mp4
```

--------------------------------

### Upgrade section example

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/github-release-notes/SKILL.md

An example of how to document breaking changes or migration steps in release notes.

```markdown
### Upgrade from 1.3
`client.run()` no longer retries by default; pass `retries: 3` to keep it.
```

--------------------------------

### Install WineCharm via Flatpak

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/winecharm/references/patterns.md

Commands to install and launch WineCharm using Flatpak.

```bash
flatpak --user install flathub io.github.fastrizwaan.WineCharm
```

```bash
flatpak run io.github.fastrizwaan.WineCharm
```

--------------------------------

### Verify WineCharm installation

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/winecharm/references/patterns.md

Command to verify the installation and view help documentation.

```bash
winecharm --help
```

--------------------------------

### App manifest configuration

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/wasm-python/SKILL.md

Example of pinning Pyodide and package versions in an application manifest.

```json
{pyodide: "0.27.x", packages: {ourpkg: "1.2.3"}}
```

--------------------------------

### Install custom wheel with micropip

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/wasm-python/SKILL.md

Use micropip to install a pure-Python package from a local file or custom index.

```python
micropip.install('file:///app/ourpkg-1.2.3-py3-none-any.whl')
```

--------------------------------

### Local LM Studio Provider Configuration

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

Example configuration for connecting to a local LM Studio instance.

```json
{
  "providers": {
    "lmstudio": { "protocol": "openai", "base_url": "http://127.0.0.1:1234/v1" }
  },
  "models": {
    "local": { "provider": "lmstudio", "model": "<model-id-as-shown-in-lm-studio>" }
  }
}
```

--------------------------------

### Build and Test Commands

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

Common Makefile targets for building, testing, and installing the PocketHarness executable.

```bash
make               # build ./pocket
make test          # functional + security + localhost curl integration tests
make sanitize      # AddressSanitizer + UndefinedBehaviorSanitizer
make install       # install to ~/.local/bin + bundled skills
make clean
```

--------------------------------

### GSAP ScrollTrigger setup and animation

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Typical setup for GSAP ScrollTrigger and an example of a scrubbed animation with pinning.

```js
gsap.registerPlugin(ScrollTrigger);
```

```js
gsap.to(".product", {
  y: -200,
  rotation: 8,
  scrollTrigger: {
    trigger: ".scene",
    start: "top top",
    end: "+=1500",
    scrub: true,
    pin: true
  }
});
```

--------------------------------

### Initialize Pyodide and Load Packages

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/wasm-python/SKILL.md

Basic setup for loading Pyodide and executing Python code. The indexURL must be an absolute path to the directory containing the Pyodide files to avoid 404 errors.

```html
<script src="https://cdn.jsdelivr.net/pyodide/v0.27.x/full/pyodide.js"></script>
<script type="module">
const { loadPyodide } = await import("https://cdn.jsdelivr.net/pyodide/v0.27.x/full/pyodide.mjs");
const pyodide = await loadPyodide({ indexURL: "https://cdn.jsdelivr.net/pyodide/v0.27.x/full/" });
// ⚠ indexURL = the directory of pyodide.js itself. Relative on a subpath = 404 hell. Absolute. Pinned.
await pyodide.loadPackage(["numpy"]);
pyodide.runPython(`print(np.array([1,2,3]).sum())`);
</script>
```

--------------------------------

### Single-GPU SFT setup and training script

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/llm-fine-tuning/references/sft-recipes.md

Sets up and runs a single-GPU SFT training job. Requires CUDA, WORLD_SIZE=1, and MODE set to 'lora', 'qlora', or 'full'. Loads a model and tokenizer, optionally applies QLoRA quantization, validates the dataset, and runs a three-step smoke test. Saves environment.json and exports the model.

```python
import inspect, json, os
from importlib.metadata import version
from pathlib import Path
import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, set_seed
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTConfig, SFTTrainer

required = {"max_length", "completion_only_loss", "eval_strategy", "save_steps"}
assert required <= set(inspect.signature(SFTConfig).parameters), "Adapt to installed TRL"
assert "processing_class" in inspect.signature(SFTTrainer).parameters
assert torch.cuda.is_available(), "This recipe requires a CUDA runtime"
assert int(os.environ.get("WORLD_SIZE", "1")) == 1, "Use a separate distributed recipe"
mode = os.environ["MODE"]
assert mode in {"lora", "qlora", "full"}
model_id, revision = os.environ["MODEL_ID"], os.environ["MODEL_REVISION"]
assert len(revision) == 40 and all(c in "0123456789abcdef" for c in revision.lower())
out = Path(os.environ["RUN_DIR"])
out.mkdir(parents=True, exist_ok=False)
seed = 1729
set_seed(seed)
bf16 = torch.cuda.is_bf16_supported()
dtype = torch.bfloat16 if bf16 else torch.float16
packages = ["torch", "transformers", "trl", "peft", "accelerate", "datasets"]
if mode == "qlora":
    packages.append("bitsandbytes")
manifest = {"versions": {p: version(p) for p in packages}, "model": model_id,
            "revision": revision, "seed": seed, "mode": mode,
            "gpu": torch.cuda.get_device_name(0), "cuda": torch.version.cuda}
(out / "environment.json").write_text(json.dumps(manifest, indent=2))
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
assert tokenizer.eos_token_id is not None, "Choose and audit the model's end token"
if tokenizer.pad_token_id is None:
    tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"
data = load_dataset("json", data_files={"train": "train.jsonl", "validation": "validation.jsonl"})
for split in data.values():
    assert len(split) and {"prompt", "completion"} <= set(split.column_names)

# FP16 AMP full tuning keeps FP32 trainable weights for gradient unscaling.
weight_dtype = torch.float32 if mode == "full" and not bf16 else dtype
load_args = {"revision": revision, "torch_dtype": weight_dtype}
if mode == "qlora":
    load_args.update(device_map={"": 0}, quantization_config=BitsAndBytesConfig(
        load_in_4bit=True, bnb_4bit_quant_type="nf4",
        bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=dtype))
model = AutoModelForCausalLM.from_pretrained(model_id, **load_args)
model.config.use_cache = False
if mode == "qlora":
    model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True)
if mode != "full":
    model = get_peft_model(model, LoraConfig(
        task_type="CAUSAL_LM", r=16, lora_alpha=32, lora_dropout=0.05,
        target_modules="all-linear", bias="none"))
    model.print_trainable_parameters()
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
assert trainable > 0

# Three optimizer steps are a mechanics smoke test, not a quality experiment.
args = SFTConfig(
    output_dir=str(out), max_length=512, packing=False,
    completion_only_loss=True, per_device_train_batch_size=1,
    per_device_eval_batch_size=1, gradient_accumulation_steps=2,
    max_steps=3, learning_rate=1e-5 if mode == "full" else 1e-4,
    optim="adamw_torch", gradient_checkpointing=True,
    bf16=bf16, fp16=not bf16, logging_steps=1,
    eval_strategy="steps", eval_steps=3, save_strategy="steps", save_steps=3,
    save_total_limit=2, report_to="none", push_to_hub=False,
    seed=seed, data_seed=seed,
)
trainer = SFTTrainer(model=model, args=args, processing_class=tokenizer,
    train_dataset=data["train"], eval_dataset=data["validation"])
batch = next(iter(trainer.get_train_dataloader()))
labels = batch["labels"]
assert (labels != -100).sum().item() > 0, "All supervised tokens were masked"
assert ((labels != -100).sum(dim=1) > 0).all(), "A row has no training target"
assert (labels[:, 1:] != -100).any(), "No target survives causal shift"
if "attention_mask" in batch:
    assert (labels[batch["attention_mask"] == 0] == -100).all()
print("Supervised tokens:", (labels != -100).sum().item())
# Inspect decoded supervised spans locally without persisting private examples.
# Audit prompt exclusion, true EOS labels and truncation before running train().
```

```python
torch.cuda.reset_peak_memory_stats()
trainer.train()
trainer.save_model(str(out / "export"))
tokenizer.save_pretrained(str(out / "export"))
print("Peak allocated bytes:", torch.cuda.max_memory_allocated())
```

--------------------------------

### Launch WineCharm GUI

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/winecharm/references/patterns.md

Command to start the WineCharm graphical user interface.

```bash
winecharm
```

--------------------------------

### Verify PocketHarness installation

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

Check that the pocket command is available and verify the version and help documentation.

```bash
command -v pocket
pocket --version
pocket --help
```

--------------------------------

### Install WineCharm dependencies on Debian

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/winecharm/references/patterns.md

List of required system packages for developer installations on Debian-based systems.

```bash
wine wine32 wine64 winetricks icoutils libimage-exiftool-perl zstd winbind python3-yaml python3-psutil libgtk-4-1 libadwaita-1-0 python3-gi
```

--------------------------------

### tshark starting patterns for traffic analysis

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/packet-trace-analysis/references/tooling-and-evidence.md

Three tshark commands for summarizing traffic mix, listing SYN packets with epoch timestamps, and following a TCP stream. Replace display filters (-Y) per question; display filters differ from capture filters which discard traffic at collection. Verify filter syntax against the installed version as field names change.

```text
tshark -r trace.pcapng -q -z io,stat,1,"COUNT(tcp) tcp","COUNT(dns) dns"
```

```text
tshark -r trace.pcapng -Y "tcp.flags.syn==1 && tcp.flags.ack==0" -T fields -e frame.number -e ip.src -e tcp.dstport -e frame.time_epoch
```

```text
tshark -r trace.pcapng -q -z follow,tcp,ascii,0
```

--------------------------------

### SVG path stroke animation setup

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Initial CSS setup for animating SVG paths by setting the stroke-dasharray and stroke-dashoffset.

```css
path {
  stroke-dasharray: 800;
  stroke-dashoffset: 800;
}
```

--------------------------------

### Convert sample rate with SoX

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/audio-processing/references/signal-pipeline.md

Illustrates a rate conversion using SoX. Confirm optional resampler availability in the installed build and consult the relevant effect help before writing a chain. Do not assume SoX is installed or that an effect exists in every build.

```bash
sox input.wav output.wav rate 48000
```

--------------------------------

### Compile and self-test C++ examples

Source: https://github.com/yunusemrejr/pocketharness/blob/main/examples/capabilities/README.md

From the repository root, compile the shortest_path and logistic_regression C++ sources with C++17 and run their self-tests. Requires a C++17 compiler.

```bash
work=$(mktemp -d)
c++ -std=c++17 -O2 -Wall -Wextra -Werror skills/algorithm-design/assets/shortest_path.cpp -o "$work/shortest-path"
"$work/shortest-path" --self-test
c++ -std=c++17 -O2 -Wall -Wextra -Werror skills/classical-ml-modeling/assets/logistic_regression.cpp -o "$work/logistic-regression"
"$work/logistic-regression" --self-test
```

--------------------------------

### Headless conversion command with unique profile

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/libreoffice-automation/references/headless-jobs.md

Example command to convert a document to PDF using a unique, isolated user profile to prevent conflicts with existing instances.

```text
soffice -env:UserInstallation=file:///tmp/office-job-UNIQUE/profile --headless --convert-to pdf --outdir /tmp/office-job-UNIQUE/output /absolute/source.docx
```

--------------------------------

### Three.js Scene Setup and Model Loading

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/threejs/assets/model-preview.html

Initializes the Three.js scene, camera, renderer, and loads a GLTF model from the document dataset.

```javascript
import * as THREE from 'three'; import {GLTFLoader} from 'three/addons/loaders/GLTFLoader.js'; const scene = new THREE.Scene(); scene.background = new THREE.Color('#121d24'); const camera = new THREE.PerspectiveCamera(45, innerWidth / innerHeight, .1, 50); camera.position.set(3.4, 2.1, 3.4); camera.lookAt(0, 0, 0); const renderer = new THREE.WebGLRenderer({antialias:true}); renderer.setPixelRatio(1); renderer.setSize(innerWidth, innerHeight); renderer.outputColorSpace = THREE.SRGBColorSpace; document.body.appendChild(renderer.domElement); scene.add(new THREE.HemisphereLight('#dbeeff', '#283126', 2)); const key = new THREE.DirectionalLight('#fff0dc', 3); key.position.set(4, 6, 3); scene.add(key); const turntable = new THREE.Group(); scene.add(turntable); let mixer; // The exporter waits for this Promise. A loader failure is propagated with its // actual source URL; an empty canvas must never become a successful final video. window.renderReady = new GLTFLoader().loadAsync(document.body.dataset.model).then(gltf => { const model = gltf.scene; const bounds = new THREE.Box3().setFromObject(model), size = bounds.getSize(new THREE.Vector3()); const span = Math.max(size.x, size.y, size.z); if (!Number.isFinite(span) || span <= 0) throw new Error('Model has no finite visible bounds'); model.position.sub(bounds.getCenter(new THREE.Vector3())); turntable.add(model); turntable.scale.setScalar(2 / span); if (gltf.animations.length) { mixer = new THREE.AnimationMixer(model); mixer.clipAction(gltf.animations[0]).play(); } window.renderFrame(0); }); window.renderFrame = seconds => { turntable.rotation.y = seconds * Math.PI / 3; if (mixer) mixer.setTime(seconds); renderer.render(scene, camera); }; addEventListener('resize', () => { camera.aspect = innerWidth / innerHeight; camera.updateProjectionMatrix(); renderer.setSize(innerWidth, innerHeight); window.renderFrame(0); });
```

--------------------------------

### Run local Colab runtime with Docker

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/google-colab-training/references/connections.md

Starts a local Docker container for Colab runtime, binding to the loopback interface.

```bash
docker run --gpus=all -p 127.0.0.1:9000:8080 us-docker.pkg.dev/colab-images/public/runtime
```

--------------------------------

### Run motion rendering script

Source: https://github.com/yunusemrejr/pocketharness/blob/main/examples/capabilities/README.md

Renders the complete six-second motion and synthesized audio demonstration. Produces 144 frames at 960x540/24 fps, checks the soundtrack, and prints encoded stream metadata. Requires Chrome and FFmpeg; the native exporter does not install them.

```sh
examples/capabilities/motion/render.sh /tmp/pocket-motion
```

--------------------------------

### Start Jupyter for Colab connection

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/google-colab-training/references/connections.md

Configures Jupyter to allow connections from the Colab origin with specific security flags.

```bash
jupyter notebook --NotebookApp.allow_origin='https://colab.research.google.com' --port=8888 --NotebookApp.port_retries=0 --NotebookApp.allow_credentials=True
```

--------------------------------

### Probabilistic Event Scheduling

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/cinematic-pixel-scene/SKILL.md

Example of using probabilistic logic to schedule events like announcements.

```js
if (time > nextAnnouncement) {
  playAnnouncement();
  nextAnnouncement = time + rand(35, 80);
}
```

--------------------------------

### Parallax Layering Implementation

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/cinematic-pixel-scene/SKILL.md

Example of separating background, midground, and foreground layers with different parallax factors.

```js
const cameraX = ...;

drawBackground(-cameraX * 0.1);
drawMidground(-cameraX * 0.4);
drawForeground(-cameraX * 0.8);
```

--------------------------------

### Navigation Structure

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Example of navigation items for long scrollytelling pages that allow jumping to scene anchors.

```text
Overview
Design
Mechanism
Performance
Details
```

--------------------------------

### Unsigned modular addition in NASM syntax

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/x86-assembly-engineering/references/patterns.md

An example of an unsigned modular addition function for System V AMD64 userspace. This implementation preserves callee-saved registers and avoids memory access.

```asm
; NASM syntax, System V AMD64 userspace: uint64_t add_u64(uint64_t a, uint64_t b)
global add_u64
section .text
add_u64:
    lea rax, [rdi + rsi]
    ret
```

--------------------------------

### State and asynchronous work: generation counter to suppress stale writes

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/browser-javascript-engineering/references/patterns.md

Use a generation counter to ignore stale responses after a newer request has started. This example suppresses stale writes but does not cancel work; add an AbortController and teardown when the component needs them.

```javascript
let generation = 0;
async function refresh(url) {
  const mine = ++generation;
  const response = await fetch(url);
  if (!response.ok) throw new Error(`HTTP ${response.status}`);
  const data = await response.json(); // validate its shape here
  if (mine !== generation) return;
  output.textContent = String(data.label);
}
```

--------------------------------

### Run tests, benchmarks, and media scripts

Source: https://github.com/yunusemrejr/pocketharness/blob/main/docs/benchmarks-0.12.0.md

Commands to run the test suite, check workflows, run benchmarks, and execute media integration and benchmark scripts. The benchmark script compares a baseline binary (e.g., 0.11.3) with the current build, and can be repeated for a specific workload. Finally, run the release test script.

```sh
make -j4 test
make check-workflows
make bench
python3 tests/media_integration.py ./pocket --out build/media-integration
python3 tests/benchmark_media.py --before /path/to/0.11.3/pocket \
  --after ./pocket --out build/media-benchmark --runs 3
# Repeat only a concerned workload:
python3 tests/benchmark_media.py --before /path/to/0.11.3/pocket \
  --after ./pocket --out build/video-confirm --runs 5 --workload video
./scripts/test-release.sh
```

--------------------------------

### Responsive Desktop Choreography Configuration

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Recommended design patterns for desktop scroll animations.

```text
large pinned scenes
horizontal movement
multi-layer parallax
side-by-side compositions
```

--------------------------------

### Build, test, and release commands

Source: https://github.com/yunusemrejr/pocketharness/blob/main/docs/execution-workflows.md

Commands to run tests, check workflows, run benchmarks, and create a tagged release. After updating kVersion and CHANGELOG and committing verified changes, tag and push the release.

```sh
make -j4 test
make check-workflows
make bench
make bench-runtime
./scripts/test-release.sh
# After updating kVersion and CHANGELOG and committing verified changes:
git tag -a vVERSION -m 'Pocket Harness VERSION'
git push origin main vVERSION
```

--------------------------------

### Execute test and benchmark commands

Source: https://github.com/yunusemrejr/pocketharness/blob/main/docs/benchmarks-0.13.0.md

Run the test suite, workflow checks, and media benchmarks using the provided make and python scripts.

```sh
make -j2 test
make check-workflows
make bench
python3 tests/media_integration.py ./pocket --out build/media-integration
python3 tests/benchmark_media.py --before /path/to/released/0.12.0/pocket \
  --after ./pocket --out build/video-pairs --runs 3 --workload video
./scripts/test-release.sh
```

--------------------------------

### DrawWindow Primitive Composition

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/cinematic-pixel-scene/SKILL.md

Example of a complex object composed of primitive rect functions.

```js
function drawWindow(x, y, w, h, opts = {}) {
  rect(x, y, w, h, opts.frame ?? COLORS.frame);
  rect(x + 1, y + 1, w - 2, h - 2, opts.glass ?? COLORS.glass);

  if (opts.glow) {
    rect(x + 2, y + 2, w - 4, h - 4, opts.glow);
  }

  if (opts.divider) {
    rect(x + Math.floor(w / 2), y + 1, 1, h - 2, COLORS.frameDark);
  }
}
```

--------------------------------

### Initialize AudioContext and Master Gain

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/cinematic-pixel-scene/SKILL.md

Sets up the base AudioContext and a master gain node to control overall volume.

```javascript
const audioCtx = new AudioContext();

const master = audioCtx.createGain();
master.gain.value = 0.7;
master.connect(audioCtx.destination);
```

--------------------------------

### Configuration File Paths

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

The locations for user and project-specific configuration files.

```text
~/.config/pocketharness/config.json
./.pocket/config.json                  # models/aliases only, never providers
```

--------------------------------

### Progressive Loading Strategy

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Load assets based on narrative progression to improve startup time.

```text
load hero immediately
load scene 2 soon
load scene 3 when scene 1 is active
load scene 4 later
```

--------------------------------

### Scene-based animation structure

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Example of dividing a page into distinct scenes for organized animation logic.

```text
Scene 1: Hero
Scene 2: Product enters
Scene 3: Product disassembles
Scene 4: Technical explanation
Scene 5: Reassembly
Scene 6: Final CTA
```

--------------------------------

### Compose recursive agency with Linux

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

Demonstrates running multiple pocket instances in parallel using standard Linux process composition.

```bash
pocket -p "Inspect auth" > /tmp/auth-review &
pocket -p "Inspect storage" > /tmp/storage-review &
wait
```

--------------------------------

### Generate and analyze procedural audio cues

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/procedural-audio/SKILL.md

Create an output directory, generate various sound effects and melodies using sfx and wav presets, and verify the output with the audio command.

```bash
mkdir -p public/audio
pocket kit sfx public/audio/reveal.wav chime --freq 660 --gain .28
pocket kit sfx public/audio/transition.wav whoosh --duration .6 --seed 17 --gain .22
pocket kit sfx public/audio/hit.wav impact --duration .4 --gain .3
pocket kit sfx public/audio/laser.wav laser --freq 1800 --end-freq 120 --duration .35
pocket kit wav public/audio/motif.wav "C5:.5 E5:.5 G5:1 R:.25 G5:.5 C6:1" --bpm 100 --wave sine --gain .22
pocket kit wav public/audio/loop.wav "C4+E4+G4:2 A3+C4+E4:2 | saw> C2:1 C2 A1 A1 | K:.5 H S H K H S H" --bpm 96 --gain .2
pocket kit audio public/audio/reveal.wav
```

--------------------------------

### Initialize and stream text generation with transformers.js

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/browser-ml/SKILL.md

Configure the pipeline to use the WebGPU device and stream generated tokens.

```javascript
import { pipeline, env } from "@huggingface/transformers";
env.allowLocalModels = false;                 // or serve from your CDN
const gen = await pipeline("text-generation", "onnx-community/Qwen2.5-0.5B-Instruct-GGUF",
  { device: "webgpu", dtype: "q4f16" });       // wasm backend auto-fallbacks
for await (const out of gen("Explain X in one sentence.",
  { max_new_tokens: 128, do_sample: false })) { /* stream */ }
```

--------------------------------

### Define subranges for scroll progress

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Example mapping of scroll progress ranges to specific animation actions.

```text
0.00–0.20   title fades
0.15–0.40   object scales
0.35–0.65   diagram appears
0.60–0.85   copy changes
0.80–1.00   scene exits
```

--------------------------------

### Produce key visuals with PocketHarness

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/key-visual-art-direction/SKILL.md

Use these commands to initialize a scene, render frames for review, and generate final video output. Edit the scene.json file to customize geometry, materials, and lighting.

```js
tool_search({ names: ["scene_create", "scene_render", "video_compose"] })
scene_create({ preset: "keyvisual" })            // frosted-glass mark, spot pool, beam, soft shadow, bloom/grain/vignette
// Edit scene.json: replace the mark with the brand's form (arch/disc/capsule/torus/box parts, group + parent),
// set backgroundGradient, wall color, spot position/target, beams, post; keep tracks for a loop.
scene_render({ path: "./media-<id>/scene.json", mode: "frame", time: 1.5 })   // LOOK at the poster, fix, repeat
scene_render({ path: "./media-<id>/scene.json", mode: "video" })              // loop or intro plate
```

--------------------------------

### Define Visual State Object

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/cinematic-pixel-scene/SKILL.md

An example of an explicit state object to avoid scattering global variables.

```js
const state = {
  time: 0,

  weather: {
    rain: 0.8,
    fog: 0.3
  },

  train: {
    x: -120,
    state: "approaching"
  },

  lights: {
    platform: true,
    vending: true
  },

  audio: {
    enabled: false
  }
};
```

--------------------------------

### Render video frames and clips with pocket kit

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/code-first-video/SKILL.md

Commands to generate a single frame, a low-resolution preview, and a final video with audio.

```bash
pocket kit frame scene.html settled.png --time 3 --size 1280x720
pocket kit video scene.html preview.mp4 --size 640x360 --fps 24 --duration 6
pocket kit video scene.html final.mp4 --size 1280x720 --fps 30 --duration 6 --audio soundtrack.wav
```

--------------------------------

### Logical Resolution Constants (480x270)

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/cinematic-pixel-scene/SKILL.md

Example of setting a low internal resolution for a cinematic pixel scene.

```js
const VIEW_W = 480;
const VIEW_H = 270;
```

--------------------------------

### Render video preview with pocket kit

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/code-first-video/references/production-pipeline.md

Command to generate a preview video with specified dimensions, frame rate, and duration.

```bash
pocket kit video scene.html preview.mp4 --size 640x360 --fps 24 --duration 6
```

--------------------------------

### Normalize scroll progress

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Helper functions to clamp values and calculate progress between specific start and end points.

```js
function clamp(v, min = 0, max = 1) {
  return Math.max(min, Math.min(max, v));
}

function progressBetween(value, start, end) {
  return clamp((value - start) / (end - start));
}
```

--------------------------------

### Compile and run shortest_path.cpp

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/algorithm-design/SKILL.md

Commands to compile the C++17 shortest path implementation and execute it with a self-test or a graph input file.

```bash
c++ -std=c++17 -O2 -Wall -Wextra -Werror assets/shortest_path.cpp -o "$TMPDIR/shortest-path"
"$TMPDIR/shortest-path" --self-test
"$TMPDIR/shortest-path" graph.txt
```

--------------------------------

### Export function with wasm-bindgen

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/wasm-rust/SKILL.md

Example of exposing a Rust function to JavaScript, mapping errors to JsError for proper exception handling.

```rust
#[wasm_bindgen]
pub fn process(input: &str) -> Result<Vec<u8>, JsError> {
    parse(input).map_err(|e| JsError::new(&e.to_string()))  // across the boundary = JS exception
}
```

--------------------------------

### Render audio cues with render.sh

Source: https://github.com/yunusemrejr/pocketharness/blob/main/examples/capabilities/audio/README.md

Executes the rendering script to generate original cues, stereo scores, and mixed audio programs in the specified output directory.

```sh
POCKET=./pocket sh examples/capabilities/audio/render.sh /tmp/pocket-audio-demo
```

--------------------------------

### Pocket Kit CLI Commands

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/video-studio/SKILL.md

A collection of commands for managing video themes, assets, narration, music, sound effects, and rendering processes.

```bash
pocket kit theme "topic words" [--seed N]          # palette + font pair derived from the subject; re-roll with --seed
pocket kit asset font "Family" fonts/ --weights 400,700   # local woff2 + css (scenes must not depend on the network)
pocket kit say --setup                              # once: Piper engine + neural voice (~90 MB); then:
pocket kit say narration.wav --file script.txt      # + narration.json/.srt/.captions.html/.cues.css (measured beat and sentence times)
pocket kit music music.wav --style corporate --duration D --seed N     # styles: ambient lofi corporate cinematic tech upbeat
pocket kit sfx whoosh.wav whoosh --seed 3 ; pocket kit sfx tick.wav click
pocket kit mix audio.wav --voice narration.wav --music music.wav --duration D --lufs -14 --at 9.3:whoosh.wav:-2 ...
pocket kit frame scene.html check.png --time 6 --size 1280x720      # lint + a still: look at it
pocket kit frame scene.html moments.png --times 3,12,40,66 --size 960x540   # several moments, ONE launch, one numbered sheet (not N renders + ffmpeg stacking)
pocket kit video scene.html preview.mp4 --duration D --fps 12 --size 640x360 --audio audio.wav
pocket kit vsheet preview.mp4 sheet.png --n 12                      # the whole video in one image: read it
pocket kit video scene.html final.mp4 --duration D --fps 30 --size 1920x1080 --audio audio.wav
pocket kit vcheck final.mp4 --duration D
```

--------------------------------

### HTML head metadata for public articles

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/search-discoverability/references/patterns.md

Example of essential SEO metadata including title, description, and canonical link for a public-facing article.

```html
<title>Calibrating an ESP32 temperature sensor | Lab Notes</title>
<meta name="description" content="A measured two-point calibration procedure, with test data and error limits.">
<link rel="canonical" href="https://example.com/lab/esp32-calibration/">
```

--------------------------------

### Read configuration with pathlib

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/python-software-engineering/references/patterns.md

Uses a context manager to safely open and parse a JSON configuration file, ensuring the result is a dictionary.

```python
from pathlib import Path
import json

def read_config(path: Path) -> dict:
    with path.open(encoding="utf-8") as stream:
        value = json.load(stream)
    if not isinstance(value, dict):
        raise ValueError("configuration must be an object")
    return value
```

--------------------------------

### Build custom package with pyodide-build

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/wasm-python/SKILL.md

Command to build a custom package for Pyodide.

```bash
pyodide build ourpkg
```

--------------------------------

### Validate audio files with FFmpeg

Source: https://github.com/yunusemrejr/pocketharness/blob/main/examples/capabilities/audio/README.md

Uses ffprobe and ffmpeg to verify the codec, sample rate, channels, and duration of generated audio files when FFmpeg is installed.

```sh
ffprobe -v error -show_entries stream=codec_name,sample_rate,channels,duration -of default=noprint_wrappers=1 /tmp/pocket-audio-demo/chime.wav
ffmpeg -v error -i /tmp/pocket-audio-demo/chime.wav -f null -
```

--------------------------------

### Spawn a child process with node:child_process

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/node-runtime-engineering/references/patterns.md

Demonstrates spawning a git process with specific working directory and stream configurations. Ensure output streams are consumed or redirected and the process tree is terminated when necessary.

```js
import { spawn } from 'node:child_process';
const child = spawn('git', ['status', '--porcelain'], {
  cwd: projectDir, stdio: ['ignore', 'pipe', 'pipe'], shell: false
});
```

--------------------------------

### Define a short musical motif in JSON

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/music-composition/references/score.md

A basic score structure defining BPM, time signature, and a single melodic track with notes containing pitch, start time, duration, and velocity.

```json
{"score":{"bpm":100,"beats":4,"numerator":4,"denominator":4,"tracks":[{"name":"Motif","program":0,"waveform":"triangle","notes":[{"pitch":60,"start":0,"duration":0.8,"velocity":86},{"pitch":64,"start":1,"duration":0.8,"velocity":80},{"pitch":67,"start":2,"duration":0.8,"velocity":90},{"pitch":64,"start":3,"duration":1,"velocity":72}]}]}}
```

--------------------------------

### Inspect workflow configurations with pocket --workflow

Source: https://github.com/yunusemrejr/pocketharness/blob/main/docs/execution-workflows.md

Use the --workflow flag to display the native complexity, reasoning, planning, and budget requirements for a specific task without initiating a session or calling a provider.

```sh
pocket --workflow 'Fix the title in README.md'
pocket --workflow 'Build a PHP 8 website and deploy through Git/SSH to Namecheap'
pocket --workflow 'Fine-tune a language model with QLoRA in Google Colab'
pocket --workflow 'Edit sound and mix music with a video animation'
```

--------------------------------

### Identify Python runtime and verify GPU availability

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/google-colab-training/SKILL.md

This standard-library cell identifies the current Python version and checks for the presence of the nvidia-smi utility to confirm GPU capability without installing additional packages.

```python
import platform, shutil, subprocess
print(platform.python_version())
gpu = shutil.which("nvidia-smi")
if gpu:
    subprocess.run([gpu, "--query-gpu=name,memory.total", "--format=csv,noheader"], check=True, timeout=10)
else:
    print("nvidia-smi unavailable; GPU capability has not been verified")
```

--------------------------------

### Render a single frame via command line

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/blender-production/references/render-pipeline.md

Executes a background render of a specific frame from a blend file, specifying output path and format.

```text
blender --background scene.blend --render-output /absolute/frames/frame_#### --render-format PNG --render-frame 1
```

--------------------------------

### Render and Audit Musical Sketches

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/music-composition/SKILL.md

Commands to render a musical sketch from a note string, check audio levels, and generate a background music bed.

```bash
pocket kit wav sketch.wav "C4+E4+G4:2 A3+C4+E4:2 | saw> C2:1 C2 A1 A1 | K:.5 H S H K H S H" --bpm 96 --gain .2
pocket kit audio sketch.wav
pocket kit music bed.wav --style ambient --duration 20 --seed 17
```

--------------------------------

### Fetch with context and HTTP client

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/go-service-engineering/references/patterns.md

Use this function to make an HTTP GET request with a context, handling cancellation and errors. It discards at most 1 MiB of the response body and returns an error for non-200 status codes. Note that it does not prove the entire body was read or that the connection is reusable; real parsing should detect oversize payloads and validate schema.

```go
func fetch(ctx context.Context, client *http.Client, url string) error {
    req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
    if err != nil { return err }
    resp, err := client.Do(req)
    if err != nil { return err }
    defer resp.Body.Close()
    if resp.StatusCode != http.StatusOK { return fmt.Errorf("unexpected HTTP status %d", resp.StatusCode) }
    _, err = io.Copy(io.Discard, io.LimitReader(resp.Body, 1<<20))
    return err
}
```

--------------------------------

### Responsive Mobile Choreography Configuration

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Recommended design patterns for mobile scroll animations.

```text
shorter pins
simpler transforms
vertical compositions
less parallax
more direct content flow
```

--------------------------------

### Run validation and benchmark commands

Source: https://github.com/yunusemrejr/pocketharness/blob/main/docs/runtime-efficiency.md

A collection of make and script commands used to verify project integrity, run benchmarks, and perform media integration testing.

```sh
make -j4 test
make -j4 test CXX=clang++  # in an independent build tree
make -j2 sanitize         # in an independent build tree
make check-workflows
make bench
make bench-runtime
./scripts/test-release.sh
python3 tests/media_integration.py ./pocket --out build/media-verification
```

--------------------------------

### LLM Iterative Development Workflow

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/cinematic-pixel-scene/SKILL.md

A 20-step loop for building scenes incrementally, ensuring each pass solves a specific class of problem.

```text
1. establish canvas and scaling
2. establish composition
3. render large silhouettes
4. inspect
5. add architecture
6. inspect
7. add palette and lighting
8. inspect
9. add props
10. inspect
11. add characters
12. inspect
13. add atmospheric effects
14. inspect
15. animate
16. inspect
17. add audio
18. inspect
19. optimize
20. final QA
```

--------------------------------

### Build a main program with em++

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/wasm-c-cpp/SKILL.md

Use em++ to compile a C++ program defining a main function into a modularized JavaScript file.

```sh
em++ main.cpp -O2 -sMODULARIZE=1 -sEXPORT_NAME=createApp \
  -sEXPORTED_FUNCTIONS='["_main"]' -o app.js
```

--------------------------------

### Project Directory Structure

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

The standard layout for source files, configuration, and data storage in PocketHarness.

```text
~/src/pocketharness/                 source / Git repo
~/.local/bin/pocket                  installed executable
~/.config/pocketharness/config.json  user config (+ roles.json, env, system.md, wisdom.md, skills/)
~/.local/share/pocketharness/        sessions, catalog.json, brain.json, bundled skills
~/.cache/pocketharness/              disposable cache
```

--------------------------------

### CONTRIBUTING.md Development and Proposal Guidelines

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/github-repo-presentation/references/patterns.md

Defines development requirements and the process for proposing changes to the repository.

```markdown
## Development
Requirements: Node >= 22, pnpm 9.
    pnpm install
    pnpm test
    pnpm lint
## Proposing a change
Open an issue for behavior changes before writing code. Keep pull requests focused;
one concern per PR. Run the checks above and paste the results in the description.
```

--------------------------------

### Execute local diagnostic commands

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

A collection of pocket kit commands for network reachability, system monitoring, SEO analysis, and asset inspection.

```sh
pocket kit reach localhost:8080 .5
pocket kit wait http://127.0.0.1:8080/health 5
pocket kit sys                  # pressure, memory and filesystem evidence
pocket kit sys 1234             # status and I/O; no environment or command line
pocket kit seo index.html       # source defects and separate recommendations
pocket kit asset inspect scene.glb
```

--------------------------------

### Run one-step value-target sanity check

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/reinforcement-learning/SKILL.md

Use this command to verify value-target calculations by providing JSON input via stdin. The script compares targets based on gamma, reward, next_value, termination, and truncation status.

```bash
python3 scripts/check_targets.py
```

```json
{"gamma":0.9,"transitions":[{"reward":1,"next_value":10,"terminated":false,"truncated":true}]}
```

--------------------------------

### Python CI

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/github-actions-workflows/references/patterns.md

Sets up a Python environment with pip caching and runs tests. It is recommended to pin the interpreter version and run linters and type checkers as separate steps.

```yaml
      - uses: actions/setup-python@v6
        with: {python-version: '3.12', cache: pip}
      - run: pip install -e '.[test]'
      - run: pytest -q
```

--------------------------------

### Sound Accessibility Configuration

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/cinematic-pixel-scene/SKILL.md

Provide a clear toggle for sound and remember user preferences locally.

```text
Sound: On / Off
```

--------------------------------

### GSAP MatchMedia Viewport Handling

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Use matchMedia to define separate timelines for different viewport widths instead of forcing one timeline to serve all.

```js
const mm = gsap.matchMedia();

mm.add("(min-width: 900px)", () => {
  // desktop timeline
});

mm.add("(max-width: 899px)", () => {
  // mobile timeline
});
```

--------------------------------

### Export frames and video with pocket kit

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/motion-graphics-production/references/rendering.md

Commands to export a single frame or a full video from an HTML scene. Specify dimensions, time, frame rate, and duration as needed.

```bash
pocket kit frame scene.html settled.png --time 3 --size 1280x720
pocket kit video scene.html output.mp4 --size 1280x720 --fps 30 --duration 6
```

--------------------------------

### Rendering technology comparison

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

A summary of use cases for DOM, SVG, Canvas, and WebGL rendering technologies.

```text
DOM:
text, layout, controls, semantic content

SVG:
diagrams, vectors, paths

Canvas:
dense 2D rendering, sequences, particles

WebGL:
true 3D, shaders, high-volume GPU effects
```

--------------------------------

### Configure Provider API Keys

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

Set environment variables for API providers. These can be set in the shell or in ~/.config/pocketharness/env.

```bash
export DEEPSEEK_API_KEY=...
export OPENROUTER_API_KEY=...     # also enables the Span/Jev judges
```

--------------------------------

### Model Configuration JSON

Source: https://github.com/yunusemrejr/pocketharness/blob/main/README.md

Defines model providers, context limits, and reasoning settings within the configuration file.

```json
{
  "models": {
    "small": {
      "provider": "ollama", "model": "qwen3:8b",
      "context": 8192, "max_tokens": 2048,
      "reasoning": "effort", "stream_usage": true
    },
    "claude": {
      "provider": "anthropic", "model": "claude-sonnet-4-6",
      "max_tokens": 8192, "reasoning": "adaptive", "prompt_cache": true
    }
  }
}
```

--------------------------------

### Configure Rain Depth Bands

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/cinematic-pixel-scene/SKILL.md

Organize rain into layers to create depth.

```text
background rain
midground rain
foreground rain
```

--------------------------------

### Content Hierarchy Planning

Source: https://github.com/yunusemrejr/pocketharness/blob/main/skills/scroll-animated-websites/SKILL.md

Identify these core elements before building animations to ensure content clarity.

```text
message
hierarchy
story order
visual emphasis
```