Claude Fable 5.2 prompting guide and community templates

Complete the entire task before yielding back.
When you encounter a recoverable obstacle (a test failure, a timeout, an empty search result, or a tool error), diagnose the cause and work around it without stopping to ask.
Respect explicit boundaries: do not touch files marked off-limits, do not bypass missing permissions, and ask before destructive or irreversible actions (such as dropping tables or pushing to production).
Every 3 to 4 tool calls, output one plain-text sentence stating what you just verified and what you are testing next.
Finish with a self-contained summary for a reader who did not watch the tool trace: what changed, how you verified it, and any open blockers.

Review this diff as the engineer responsible for production reliability.
Do not stop at syntax or whether existing unit tests pass. Trace every modified function across its callers, check state transitions, race conditions, partial failures, and unit mismatches, and verify what happens when upstream inputs are empty, stale, or out of range.
List only concrete bugs with the exact file, line range, failure trigger, and minimal fix.

Create a [10]-second, [1920x1080, 30fps] 3D animation of [subject and action] by driving Blender headless from the command line with Python.
Build all meshes, rigging, skin weights, materials, lighting, and camera motion procedurally in Python—do not download external models or textures.
Workflow:
1. Build the scene and render 4 low-resolution keyframe PNGs (front, side, 3/4, camera view).
2. Inspect the rendered PNGs for mesh clipping, broken proportions, flat lighting, or dead framing. Fix every issue you see in the script and re-render the test frames until they hold up.
3. Log the render time per frame, render the full sequence, and encode to MP4. Deliver both the final MP4 and the Python script.

Build a single self-contained HTML file using Three.js that renders a [14]-second looping 3D scene at [1280x720, 30fps]: [describe scene, lighting, and motion already active at t=0].
Requirements:
- Generate all geometry, shaders, particles, and materials in code.
- Make the animation strictly deterministic and expose window.renderAt(seconds) so any timestamp renders the exact frame.
- Open the page in a headless browser, capture screenshots at t = 0s, 3s, 7s, and 12s, inspect them for visual bugs or frame-rate bottlenecks, and refine the scene before encoding the final MP4.

Build a polished, playable browser game in a single HTML file: [genre, one arena/level, core movement and scoring rules].
- Generate all 3D/2D geometry, procedural textures, particle effects, and WebAudio sound effects in code with no external assets.
- Expose a debug state hook and run an automated headless playtest that simulates player inputs, checks for console errors, verifies collision and scoring, and captures screenshots of gameplay.
- Fix any broken physics, unreadable UI, or frame-time spikes before delivering the final file.

Step 1: Define a fictional painter—era, subject matter, palette, brush physics, and how they layer coarse underpainting vs fine detail strokes.
Step 2: Write a standalone Python program that paints [subject] in that artist's style on a [2000x1500] canvas. Simulate bristle brushes, pigment mixing, canvas tooth, and stroke direction; place every stroke algorithmically with no image generation APIs and no source image tracing.
Step 3: Run the script to output a PNG, inspect the image against the painter's style, and iterate on the brush engine, composition, and lighting for at least 3 passes before finishing.

curl https://api.keelrouter.com/v1/messages \
  -H "x-api-key: $KEELROUTER_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-fable-5-2",
    "max_tokens": 32000,
    "output_config": { "effort": "high" },
    "system": "Complete the entire task before yielding back. Diagnose recoverable tool or test errors yourself and give a one-line status update every few steps.",
    "messages": [
      {
        "role": "user",
        "content": "Audit src/lib/proxy/chat.ts for stream settlement edge cases and write a reproduction test."
      }
    ]
  }'