Hatch Pet

Hatch Pet (Claude Edition)

Overview

Produce a Claude-compatible animated pet spritesheet and pet.json. The user generates each image manually via the GPT Image 2 web interface (chatgpt.com) using the prompts Claude prepares. Claude handles all deterministic work: folder preparation, prompt generation, image processing, atlas assembly, QA, and packaging.

Image generation is manual: Claude never calls any image API. Instead, Claude outputs each prompt clearly, the user copies it into ChatGPT web, and then provides the generated image file to Claude.

Visible Progress Plan

Keep this checklist visible for every pet run. Update it as each step completes.

1. Getting <Pet> ready.
2. Imagining <Pet>'s main look. (user generates base image)
3. Picturing <Pet>'s poses. (user generates 9 animation rows)
4. Hatching <Pet>. (Claude processes, validates, and packages)

Only mark a step done when the real file or decision exists.

Generation Prompt Format

When presenting a prompt for the user to paste into ChatGPT:

1. Show the image type (Base Character / Row: idle / Row: running-right / etc.)
2. Show the frame count and size expected
3. Show the full prompt text in a code block
4. List the GPT Image 2 settings to use
5. Tell the user what to name/save the file as and where to put it

After showing a prompt, wait for the user to confirm they have the image before moving on.

Default Workflow

Step 1 — Prepare the run

SKILL_DIR="${CLAUDE_HOME:-$HOME/.claude}/skills/hatch-pet"
python "$SKILL_DIR/scripts/prepare_pet_run.py" \
  --pet-name "<Name>" \
  --description "<one sentence>" \
  --output-dir /absolute/path/to/run \
  --pet-notes "<stable pet description>" \
  --style-preset auto \
  --force

Optional flags:

• --reference /path/to/ref.png — user-supplied reference art
• --style-notes "<freeform>" — extra style guidance

This creates:

run/
  pet_request.json
  imagegen-jobs.json
  prompts/base-pet.md
  prompts/rows/<state>.md       (9 files)
  prompts/row-retries/<state>.md
  references/layout-guides/<state>.png

Step 2 — Base character

Read prompts/base-pet.md. Present it to the user:

─────────────────────────────────────────────
BASE CHARACTER — paste this into ChatGPT (GPT Image 2):
─────────────────────────────────────────────
<full prompt text>
─────────────────────────────────────────────
Settings: Size 1024×1024, Quality: high
Save as: <run-dir>/decoded/base.png
─────────────────────────────────────────────

Wait for the user to provide the file. Once received:

• Copy it to decoded/base.png
• Copy it to references/canonical-base.png (this becomes the identity lock for all rows)
• Mark the base job complete in imagegen-jobs.json

Step 3 — Animation rows

Generate rows in this order:

1. idle
2. running-right
3. running-left (may be mirrored from running-right — check first)
4. waving
5. jumping
6. failed
7. waiting
8. running
9. review

For each row, read prompts/rows/<state>.md and present it to the user:

─────────────────────────────────────────────
ROW: <state> (<N> frames) — paste this into ChatGPT:
─────────────────────────────────────────────
<full prompt text>
─────────────────────────────────────────────
Settings: Size 1792×1024 (wide), Quality: high
IMPORTANT: Also attach your canonical-base.png as a reference image.
Save as: <run-dir>/decoded/<state>.png
─────────────────────────────────────────────

Wait for the user to provide the file. Copy to decoded/<state>.png and mark the job complete.

running-left mirror check: After receiving running-right, show the image to the user and ask:

"Does this pet look safe to mirror for running-left? Check that no asymmetric markings, props, or text would look wrong when flipped." Only mirror if the user confirms. Otherwise generate normally.

Step 4 — Mirror running-left (if approved)

python "$SKILL_DIR/scripts/derive_running_left_from_running_right.py" \
  --run-dir /absolute/path/to/run \
  --confirm-appropriate-mirror \
  --decision-note "<user confirmation note>"

This mirrors each frame individually, preserving frame order and timing.

Step 5 — Process, validate, and QA

Run all deterministic scripts after every row is received and placed:

RUN_DIR=/absolute/path/to/run
SKILL_DIR="${CLAUDE_HOME:-$HOME/.claude}/skills/hatch-pet"
mkdir -p "$RUN_DIR/final" "$RUN_DIR/qa"

python "$SKILL_DIR/scripts/extract_strip_frames.py" \
  --decoded-dir "$RUN_DIR/decoded" \
  --output-dir "$RUN_DIR/frames" \
  --states all \
  --method auto

python "$SKILL_DIR/scripts/inspect_frames.py" \
  --frames-root "$RUN_DIR/frames" \
  --json-out "$RUN_DIR/qa/review.json" \
  --require-components

python "$SKILL_DIR/scripts/compose_atlas.py" \
  --frames-root "$RUN_DIR/frames" \
  --output "$RUN_DIR/final/spritesheet.png" \
  --webp-output "$RUN_DIR/final/spritesheet.webp"

python "$SKILL_DIR/scripts/validate_atlas.py" \
  "$RUN_DIR/final/spritesheet.webp" \
  --json-out "$RUN_DIR/final/validation.json"

python "$SKILL_DIR/scripts/make_contact_sheet.py" \
  "$RUN_DIR/final/spritesheet.webp" \
  --output "$RUN_DIR/qa/contact-sheet.png"

python "$SKILL_DIR/scripts/render_animation_previews.py" \
  --frames-root "$RUN_DIR/frames" \
  --output-dir "$RUN_DIR/qa/previews"

If preview GIFs show extraction-induced size popping (source strip itself was stable), rerun with --method stable-slots:

python "$SKILL_DIR/scripts/extract_strip_frames.py" \
  --decoded-dir "$RUN_DIR/decoded" \
  --output-dir "$RUN_DIR/frames" \
  --states all \
  --method stable-slots

python "$SKILL_DIR/scripts/inspect_frames.py" \
  --frames-root "$RUN_DIR/frames" \
  --json-out "$RUN_DIR/qa/review.json" \
  --require-components \
  --allow-stable-slots

Step 6 — Visual QA

Show the user qa/contact-sheet.png and the GIFs in qa/previews/. Ask:

"Check all 9 rows. Do any rows have identity drift, clipped bodies, wrong motion, or visible artifacts?"

If QA passes, continue to packaging.

If a row fails, show the user the specific failure note from qa/review.json, then re-present the row's prompt with a repair note appended. Once the user provides a replacement image, overwrite decoded/<state>.png and re-run Step 5.

Step 7 — Package

RUN_DIR=/absolute/path/to/run
PET_ID=$(jq -r '.pet_id' "$RUN_DIR/pet_request.json")
DISPLAY_NAME=$(jq -r '.display_name' "$RUN_DIR/pet_request.json")
DESCRIPTION=$(jq -r '.description' "$RUN_DIR/pet_request.json")
PET_DIR="${CLAUDE_HOME:-$HOME/.claude}/pets/$PET_ID"
mkdir -p "$PET_DIR"
cp "$RUN_DIR/final/spritesheet.webp" "$PET_DIR/spritesheet.webp"
jq -n \
  --arg id "$PET_ID" \
  --arg displayName "$DISPLAY_NAME" \
  --arg description "$DESCRIPTION" \
  '{id: $id, displayName: $displayName, description: $description, spritesheetPath: "spritesheet.webp"}' \
  > "$PET_DIR/pet.json"

jq -n \
  --arg run_dir "$RUN_DIR" \
  --arg package "$PET_DIR" \
  '{ok: true, run_dir: $run_dir, package: $package}' \
  > "$RUN_DIR/qa/run-summary.json"

Report the output paths to the user.

Image Requirements Summary

When the user asks what images to provide, show this table:

Total: 10 images (1 base + 9 rows, or 9 if running-left is mirrored).

Always instruct the user to attach the base character image as a visual reference when generating each row.

Prompt Requirements for the User

Every prompt must produce:

• A flat solid chroma-key background (default: #00ff00, unless it conflicts with pet colors)
• No text, logos, scenery, UI, shadows, reflections, or floor
• Clean silhouette readable at 192×208 px
• For row strips: frames laid out horizontally side by side, evenly spaced, no borders or guides

Running-Left Mirror Rule

Only mirror running-right when ALL of the following are true:

• No asymmetric markings or prop placement that reads wrong when flipped
• The prop still makes sense on the mirrored side
• Frame timing stays in the same order
• Direction clearly becomes leftward

If the user has any doubt, generate a separate running-left.

Repair Workflow

If a row fails QA:

1. Show the user the failure note from qa/review.json
2. Re-present the row's original prompt with a compact repair note:

   REPAIR NOTE: <specific issue from review.json>
   Keep everything else identical to the previous generation.
   

3. Wait for replacement image
4. Overwrite decoded/<state>.png
5. Rerun Step 5 for the full pipeline

For extraction-induced motion popping: do not ask the user for a new image first. Try --method stable-slots on the existing decoded file.

Job Manifest Updates

After copying each received image to its decoded path, mark it complete:

JOB_ID=<state>
SOURCE=/absolute/path/to/decoded/<state>.png
UPDATED_AT=$(date -u +%Y-%m-%dT%H:%M:%SZ)
TMP=$(mktemp)
jq --arg id "$JOB_ID" --arg source "$SOURCE" --arg at "$UPDATED_AT" \
  '(.jobs[] | select(.id == $id)) += {status: "complete", source_path: $source, completed_at: $at}' \
  "$RUN_DIR/imagegen-jobs.json" > "$TMP"
mv "$TMP" "$RUN_DIR/imagegen-jobs.json"

Acceptance Criteria

• final/spritesheet.webp is exactly 1536×1872
• 8 columns × 9 rows, 192×208 px per cell
• Unused cells are fully transparent
• qa/review.json has no errors
• Contact sheet and GIF previews reviewed and accepted by user
• ~/.claude/pets/<pet-id>/pet.json and spritesheet.webp written

Rules

• Never call any image API or generate images programmatically.
• Always wait for the user to provide each image before proceeding.
• Always show the full prompt text for the user to copy.
• Always instruct the user to attach the canonical-base.png when generating rows.
• Only mirror running-left with explicit user confirmation.
• Repair the smallest failing scope first (one row, not the whole sheet).
• Keep the pet's silhouette, palette, style, and props consistent across all rows — treat identity drift as a blocker even if scripts report no errors.