Generate the Son-of-Anton (SoA, Tier 4) sprite atlas for an EXISTING Codogotchi pet that has a Codex `spritesheet.webp`. Produces the 10-row `codogotchi-soa-spritesheet.webp` animating delivery-gate moments (celebrating, hyped, reviewing, pushing, etc.). Needs only the Codex sheet (independent of the Lite tiers). Use when adding SoA delivery-gate reactions to a pet.
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Generate the Son-of-Anton (SoA, Tier 4) sprite atlas for an EXISTING Codogotchi pet that has a Codex `spritesheet.webp`. Produces the 10-row `codogotchi-soa-spritesheet.webp` animating delivery-gate moments (celebrating, hyped, reviewing, pushing, etc.). Needs only the Codex sheet (independent of the Lite tiers). Use when adding SoA delivery-gate reactions to a pet.
Paths in this skill — scripts/…, references/…, and README.md below are relative to this plugin's root (hatch-codogotchi/, two directories up from this file). cd to the plugin root before running the commands, or prefix each path with it.
I/O policy — scripts own all paths; the skill names none.image_gen cannot reliably store or re-find what it generates, so never tell it to save a row to a named directory and never hunt for an image-generation save folder. Pass whatever local/temp paths the caller picks into the script arguments and chain each script's output forward; the only artifact that must persist is the final magenta atlas the user keys and installs.
No model-side visual inspection. Do not use Computer Use, screenshots, or UI automation to eyeball generated output — the model cannot reliably see its own framing and it burns tokens. Frame geometry is enforced deterministically by slice_grid.py; visual review is the human's, on the script-produced contact sheet and previews.
hatch-codogotchi-soa
Generate the Tier 4 (SoA) sprite sheet for an existing Codogotchi pet — the 10-row codogotchi-soa-spritesheet.webp that animates Son-of-Anton delivery gate moments. (SoA was Tier 3 before the Lite sheet split into Basic + Enhanced.)
Prerequisite: the pet must already have a valid spritesheet.webp (Codex, Tier 1) installed. The character reference is derived directly from that sheet — no separate seed image or description is needed. The SoA sheet needs only the Codex sheet — it is independent of the Lite tiers (Basic/Enhanced).
Execution model (slot-first): Codex uses its built-in image_gen tool to generate one full SoA row per call as 8 invisible equal-size slots, 4 across and 2 down (read left-to-right, top row then bottom), all 8 filled, on the single flat chroma key chosen for this pet (see below). The slots are an invisible placement grid — the prompt says "slots," never "grid/cells," so the model paints one continuous flat field, not 8 separate cards. prepare_pet_run.py also emits a layout-guide.png (4×2 placeholder with slot boundaries, blue safe-margins, centering crosshairs); attach it (and the seed) to every image_gen call as reference-only — it keeps frames from clipping. image_gen need not hit an exact canvas size. Then run slice_grid.py (dimension-tolerant): it slices by fraction, shares one scale across all 8 frames, and emits the exact canonical 1536×208 row strip. Generate one row at a time; never ask for the whole atlas in a single image.
Non-negotiable row gate: finish one row completely (generate grid → slice_grid.py) before generating the next. If a grid comes back clipped, cramped, or off-model, regenerate the whole grid instead of patching forward. Do not compose until every row strip exists. The model does not screenshot or eyeball its own output — slice_grid.py enforces geometry, and prop clarity / face / scale / motion are reviewed by the human on the script-produced contact sheet after composing.
Chroma key — one auto-selected key for the whole pet, keyed by the user later. The chroma is fixed for the pet by prepare_pet_run.py (auto-selected from the seed — magenta #FF00FF, blue #0047BB, or green #00B140 — magenta fallback) and must match the pet's existing sheets; pass --chroma to match if needed. Reserved success accents recolor automatically to avoid the key: green-tdd's and review-clean's ✓ effects default to green but flip to blue under a green key, so they are always preserved, never keyed out. This plugin does not key the sheet — the user keys it later in Codogotchi Studio (https://codogotchi.app/studio).
Recommended production pattern: generate the minimum number of distinct keyframes needed for a readable, non-static row, then reuse or mirror earlier stable frames to close the loop when that preserves the emotional beat. Many SoA rows can be finished faster with ~4 strong keyframes plus a mirrored/reused closure instead of 8 fully independent generations. This is a preferred optimization, not a universal law.
These rows are only shown when ~/.codogotchi/gate.json is active and unexpired. Hooks must be installed (Settings → General) and the Son-of-Anton delivery tool must emit the gate names. The sheet alone does not trigger SoA animations.
Character reference — extracted from the existing Codex sheet
The existing spritesheet.webp defines the character. Extract a reference cell before generating any frames:
This extracts the idle row, frame 1 (row 0, col 0) on a solid #FF00FF background. Inspect the seed artifact; if the pose is unclear, pass --row / --col to pick a better cell. Attach this image to every strip generation call — the SoA sheet must be indistinguishable in style from the existing Codex sheet. Every SoA row is generated on flat #FF00FF magenta and keyed by the user later; green props are preserved.
Critical failure modes — read before generating a single frame
Faking frames by transforming the seed. Each frame must be a genuine image-generation render of the character in that distinct pose. Code (Pillow) is post-processing only.
Rushing the whole atlas in one pass. One row at a time, to completion. Whole-atlas generation is a shortcut = reject.
Clipped frames. The generation format is one 4×2 grid of 8 cells, each exactly 192 × 208, no empty cell. The 4×2 framing is specifically what keeps the character from clipping the cell edge — never request a single wide 8×1 strip. If any foreground crosses a cell boundary, regenerate the grid. Generate one grid per row; never the whole atlas at once.
Style drift from the Codex sheet. After each row, compare a frame side-by-side with a Codex cell. Regenerate if the style, palette, or proportions have shifted.
Visual identity drift. Every frame must preserve the same age/proportions, hair silhouette, outfit/accessories, palette, and linework as the seed artifact. This is an eyeball pass on the contact sheet — there is no automated identity gate.
Jerky / over-animated motion (the stability killer). Ask for big or whole-body motion and it comes back incoherent across the row — legs swing, props teleport, the pet hops. Stability beats expressiveness even on these celebration rows: anchor the body and both feet, move one element at low amplitude, keep frame-to-frame change small. Sell the beat with pose and face, not with the body roaming the cell. A mild stable loop beats a busy jittery one. The scripts cannot detect this — it is purely an eyeball check.
Validation does not catch failures 1–2 or 6. The human reviews every finished row on the contact sheet for genuine but stable motion — the model does not screenshot or eyeball its own output.
Mirrored/reused closure is allowed only if it still reads as lively and intentional. If the row's reaction looks flattened or repetitive, add a distinct frame — but never buy expressiveness with jitter or limb flailing; a calm readable beat wins.
Standing constraints (every frame, every row)
Identical to hatch-codogotchi-lite:
Background: the pet's single flat chroma key on every row (use the exact key from the generated prompt / run-config.json — green-tdd and review-clean keep their ✓ accents, which recolor to stay off the key, preserved and keyed by the user later) — do NOT request RGBA directly. The key must be perfectly flat: no lighting falloff, vignette, texture, shadow, halo, glow, or antialias spill into the background.
Padding: ≥ 8 px all sides; nothing touches an edge.
Scale registration: one shared scale per row, applied by slice_grid.py (tallest cell sets it).
Horizontal registration: character/content stays on a stable x-axis in every 192×208 cell; no left/right hopping. If a large side prop skews the alpha bbox, prefer the character body's visual center and confirm by human review.
Baseline registration: feet on same y-line — baseline_y = 208 − 8 − scaled_h. Do not vertically center ordinary standing rows; they should sit near the bottom of the cell. Explicit jump/leap rows may leave the baseline briefly but must visibly take off and land.
Motion restraint (paramount): stability beats expressiveness. Keep the torso, head, hips, and both feet anchored in nearly the same place across the row and confine motion to one element (the prop, one arm, the expression) at low amplitude with short smooth arcs. Legs do not swing or restage between frames. Props travel a little and consistently — never roaming around the cell.
Loop closure: frame 8 pose ≈ frame 1 pose.
Character fidelity: seed image is sole style reference.
Green is fine on props: intentional green details are preserved for the user's keying tool; only keep the background key perfectly flat.
SoA rows are expressive — these are delivery gate reactions, not idle loops — but expressiveness comes from a clear pose and face, not from the body roaming the cell or limbs flailing. Each row should read as a distinct emotional beat at a glance while staying stable frame-to-frame. The only sanctioned big motion is ticket-completed's leap (feet off baseline ≤ 12 px, with a clean takeoff and landing on a stable x-axis); every other row sells its emotion from a planted stance. When expressiveness and stability conflict, choose stability.
Workflow
Step 1 — Extract the character reference from the existing Codex sheet
Inspect the seed artifact. The character should be in a clean neutral pose on #FF00FF. If the idle frame is unclear, try --row 3 --col 0 (waving) or another expressive cell.
# Check cell dimensions if not standard 192×208
python scripts/extract_seed_from_codex.py \
--spritesheet "${CODOGOTCHI_HOME:-$HOME/.codogotchi}/pets/<pet-id>/spritesheet.webp" \
--print-cell-size
Step 2 — Prepare the run
python scripts/prepare_pet_run.py \
--seed <seed> \
--pet-name "<existing pet display name>" \
--pet-id "<existing pet id>" \
--tier soa \
--style auto
Produces a run manifest and strip prompt artifacts under the working directory chosen via --out-dir. Use those artifacts as inputs to the path-agnostic pipeline below.
Step 3 — Generate row strips (one row at a time)
For each of the 10 SoA rows, in the order below, complete the full cycle before starting the next:
Read motion description in sheet-prompts/soa/<row-label>.txt.
Use built-in image_gen to generate one row of 8 invisible 4×2 slots — 4 across, 2 down, all 8 filled, no empty slot, on the pet's single flat chroma key (from the prompt / run-config.json). image_gen need not hit an exact size. green-tdd and review-clean keep their ✓ accents (recolored to stay off the key, preserved, keyed by the user later). Attach the seed artifact and layout-guide.png as references. The generated image is an ephemeral input to slice_grid.py — do not save it to a named directory.
Compare style to a cell from the existing spritesheet.webp — palette, linework, and proportions must match.
Step 4 — Slice each grid into its canonical row strip
slice_grid.py is dimension-tolerant (any input size), enforces 192×208 cell geometry, shares one scale across the row, and emits the exact 1536×208 strip. Paths are caller-chosen/ephemeral — the skill names none. Prop clarity, scale, and identity are reviewed by the human on the contact sheet after composing, not by the model eyeballing its output.
If any cell in a grid is wrong, regenerate the whole grid for that row, re-run slice_grid.py, and re-run the QA scripts. Fix by regenerating the grid, never by editing individual cells.
python scripts/make_contact_sheet.py \
--atlas <atlas-webp> --tier soa
python scripts/render_animation_previews.py \
--atlas <atlas-webp> --tier soa
python scripts/pre_install_qa_gate.py \
--atlas <atlas-webp> --tier soa
Step 8 — QA style cross-check
Open the SoA contact sheet alongside the existing Codex contact sheet. The character must read as the same individual — same proportions, palette, and linework — despite the more energetic poses. Regenerate any row that looks like a different character.
Step 9 — Key, then install alongside existing pet
The composed codogotchi-soa-spritesheet.webp still has its flat magenta background. Do not install it directly. Direct the user to https://codogotchi.app/studio to key it (load → tune tolerance/edge/spill → export the transparent sheet). Install the keyed SoA sheet beside the existing pet; do not overwrite spritesheet.webp, Lite sheets, or pet.json.
Quit and reopen Codogotchi or re-select the pet in Settings.
SoA row generation order
Generate rows in this order. Complete each row fully before starting the next.
See references/animation-rows-soa.md for full motion descriptions.
Order
Row index
Label
gate / activity_state
Moment
1
0
ticket-started
ticket_started
Ticket work begins
2
1
red-tdd
red_tdd
Failing test recorded (expected)
3
2
green-tdd
green_tdd
Test now passes
4
3
adversarial-review
adversarial_review
Calling for adversarial reviewer
5
4
open-pr
open_pr
PR opened and presented
6
5
poll-review
poll_review
Awaiting AI review
7
6
review-clean
review_clean
Review came back clean
8
7
record-review
record_review
Logging the review outcome
9
8
advance
advance
Moving the stack forward
10
9
ticket-completed
ticket_completed
Ticket complete — jubilant
Emotional arc: the row sequence tells a story — pumped start → careful TDD → relief at green → calling backup → proud PR → patient wait → delight at clean review → diligent logging → forward momentum → full celebration. Each row must read as its own distinct beat, not a variation of the same pose.
Key distinctions to preserve:
red_tdd is determined, not sad (failing test was intentional)
poll_review is calm anticipation, not panic
review_clean is lighter relief than ticket_completed's big jump
record_review is diligent note-taking, not celebratory
ticket_completed is the most jubilant row — full leap with both feet off baseline (≤ 12 px)
Flat #FF00FF background, perfectly uniform across the atlas (no falloff/shadow/texture)
Character/content horizontal center is stable across each row; non-jump poses share a bottom foot baseline near cell_h - 8 (eyeballed)
No row has all 8 frames pixel-identical (gated by validate_atlas.py)
Stable motion (paramount): body/feet anchored, one element moves at low amplitude, no jitter/hopping/limb-swing; emotion sold from a planted stance (only ticket-completed leaves the baseline)
Each row shows subtle distinct motion (a floor, not big motion) and reads as its named emotional beat (eyeball check)
Per-frame visual QA passed: same age/proportions, hair silhouette, outfit/accessories, palette, and linework as the seed artifact
Loop closes: frame 8 flows back to frame 1
All 10 rows have meaningfully distinct visual language from each other
validate-soa.json, contact-soa.png, and previews-soa/ exist and are newer than the final atlas
pre_install_qa_gate.py passed; user directed to https://codogotchi.app/studio to key the magenta atlas before install
After keying, installed as codogotchi-soa-spritesheet.webp beside existing spritesheet.webp; app shows SoA animations when gate.json is active (requires hooks installed)
Final response checklist
Before saying done, report: rows generated or regenerated; validation command/result; contact sheet and preview directory paths; known compromises. State clearly that the delivered atlas is magenta-background (pre-key) and the user must key it at https://codogotchi.app/studio before installing. If the tier was completed unusually quickly, state what was compressed, reused, or skipped. Validation alone is not QA.
Timing note: review_clean vs record_review
On the happy path, SoA fires review_clean then record_review seconds apart — record_review overwrites review_clean in gate.json. So review_clean's loop may only flash briefly. The distinct art is still worth it; don't skip it in generation.
