Not read at runtime. How the skill was built and why it makes its choices.
\nDistilled on 2026-09-30 from a browser side-scroller built over two days in one agent session: one hero, two ground enemies, five levels of one setting. Every generation ran through the Scenario MCP server (recommend, model_schema_get, model_run with dry_run and wait: false, jobs_wait, upload_asset and upload_asset_complete, asset_download); keying, loop search, pixelation, seamless tiles, prop cutting and WebP conversion ran as local Python. The sources were the build journal, the verbatim prompt record and the build's own processing scripts.
The hero route is an earlier isometric sprite-cycle pipeline (two-pose still on flat magenta, reframe, image-to-video clips, loop-window search, one shared palette) adapted to a single side-view facing. The build used a text-to-image model for the stills, the pose sheet and the level layers; an image-to-video model with a last-frame input for the clips; and a pixel-art model for a superseded first pass. The skill names no model and discovers both lanes with recommend (txt2img, and img2video with features: [\"endImage\"] for pinned clips), because availability differs per team and these generations will be superseded. The record has no credit costs, which is why the body prices with dry_run instead of quoting a figure.
The build also scored every level and gave each action a sound effect. One skill, one output type: the deliverables here are images (sprite strips and level layers; the clips are intermediates), so music and effects stay with scenario-audio and scenario-elevenlabs. The body says in one sentence that they are outside this skill without naming those siblings: a named sibling joins the context fan-out and the install commands, and the image workflow never calls them. The prompt templates and processing settings carry no audio sections for the same reason.
Keying, the loop search, the shared-canvas palette and the seamless cross-fade all need the pixels on disk, and the deliverables are game files, not platform assets. scenario-sprite-animation lists \"assembling locally\" as a mistake for sheets the platform can tile itself; this skill's local steps are the ones no MCP tool performs, so the body says why once and offers upload_asset into a collection for users who want the finished strips on the platform.
--drop reproduces that, applied after the canvas and palette, which is what matches the shipped strips.process_env.py rather than trusting the prompt.All four run in plain Python 3 with Pillow and NumPy; side_sprites.py also needs ffmpeg.
reframe_stills.py: from the isometric pipeline, pixels unchanged. The output names lost their isometric facings: it now writes <name>_rest.png (left pose, the first frame of every ground clip) and <name>_action.png (right pose: hero jump or enemy attack wind-up). The build ran its own copy with an auto split column and a death cycle; the shared copy reproduced all six of the build's first frames pixel for pixel, so those extras were not ported.sprite_cycles.py: the isometric pipeline's processing core, now a helper module imported by side_sprites.py only. Its four-facing command line and anything only that command line used were removed. Two changes after the port's review: pick takes the cycle's mode and windows by it (it used to branch on the name attack, so a loop named attack got one-shot windowing), and read_frames stops on a clip more than 2% off square instead of squashing it, since the figure scale and feet anchor assume a square frame. Square clips process as before.side_sprites.py: the build's one-facing wrapper, generalized. The hard-coded hero cycle table, pose names and file names became --cycles, --pose-names and <name>_pose_*; loop cycles other than walk, run and idle need a --range; --drop was added for the frames removed by hand; missing clips are an error instead of a silent skip; overwrites need --force; --webp writes lossless WebP; sprite_cycles.py is imported from the script's own folder.process_env.py: the build's script, generalized. Hard-coded input and output folders and the fixed five-level loop became --env, --out, --levels and --file KIND=PATH; texture size, fade share, ledge height, prop count and height, alpha threshold and the three WebP qualities became flags; --dry-run and --force were added. The processing is unchanged.Comments, docstrings and help text were moved to American spelling and house style in the port; no processing changed.
\nRun on 2026-09-30 (Windows, Python 3.14.2, NumPy 2.5.0, Pillow 12.2.0, ffmpeg 8.1.1) against the build's intermediate files, writing only to a temporary folder, with the scripts as they stood before the port's renames:
\n--help printed usage for every script.reframe_stills.py on the hero still and both enemy stills: all six first frames pixel-identical to the build's.side_sprites.py for the hero (eight cycles, the pose sheet, the jump still, --drop atk1=2), the heavy enemy (--drop attack=5) and the armed enemy (--range walk=12:48): all 19 strips and poses pixel-identical to the build's. Cells, anchors and fps matched the build's meta files: hero 98x81 anchored at 49,75; heavy enemy 137x123 at 68,111; armed enemy 116x93 at 58,92. Seams printed: hero run 0.32, idle 0.282, heavy-enemy walk 1.032 (flagged), armed-enemy walk 0.421, the first armed-enemy walk 1.445 (flagged). Without --drop the hero's first attack has 7 frames and the heavy enemy's attack 8, as recorded.side_sprites.py refused to overwrite without --force, stopped on a missing clip, and asked for --range on an unknown loop cycle; --webp wrote lossless WebP.process_env.py over five levels: all 35 outputs byte-identical to the build's WebP layers. A second run refused to overwrite; --file tex=... --file props=... wrote env_tex_<stem>.webp and env_prop_<stem>_<k>.webp; with neither --env nor --file it stopped with an error. On the five textures the mean wrap-edge difference after the cross-fade sat at the neighboring-pixel level.Apart from the two sprite_cycles.py fixes above, the port changes names and text only. tests/scenario-side-view-game-kit/ runs on synthetic inputs: the build's own files are not in this repository, so the byte-level reproduction above is not re-run in CI.
references/prompt-templates.md gives a template in the style of the other prompts and says so. The attack, lunge, shot, hurt and spin motion sentences are abbreviated in the record; the templates keep their recorded opening and closing words.