fuxiaoji/ai-wargame-video-skill

AI Wargame Video

Connect what the model knew at the time, the orders it issued, and what the rules actually resolved into a story the audience can understand. Reuse the user's existing games, production packages, and editing projects; this skill does not include a game engine, rulebook, model service, or third-party assets.

First Determine What to Do This Round

When the user only modifies the cover, narration, initial setup, or editing, reuse completed games and assets and only update the relevant deliverable. Only start a new batch when the user explicitly requests a replay. If a tool is missing, continue with work that doesn't depend on it; do not equate generating JSON, extracting frames, or saving drafts with native opening and exporting.

First read the project governance and production package's entry points, data interfaces, asset sources, and acceptance records; verify the latest game selection rather than reusing old scenarios, old folders, or automatic rankings. Record the engine/rule version, actual model identifier, seed, formation, perspective, assets, and output directory. Do not export account information, API credentials, private reasoning, or the entire working directory.

Read the reference based on the current task:

• New game, replay, friend-or-foe identification, radar, weather, or telegrams: Double-Blind and Sensors.
• Batch games, supervision, game selection, recovery, asset archiving, and final delivery: Production Process.
• Historical background, dialogue, interaction, imitation editing, AI image, or cover: Narrative and Visual.
• Running attached tools, integrating existing engines, or timelines: Lightweight Data Contract.

Main Pipeline from Game Facts to Finished Video

1. Freeze the playable setup. Divide according to the user's formations; check for overlapping spawns, land, boundaries, distance, and heading. List the original source, implemented rules, and current expansion separately. Implement and validate expansions before freezing; rain and lighting effects cannot replace rule adjudication.
2. Independent command for each formation. Each seat has an independent session, memory, and its own observations; friend initial setup is provided as a static time-stamped brief. Other formations can only influence it through actually delivered communications. The director's full view does not flow back to the model.
3. Issue real orders by phase. Record inputs, accepted orders, public tactical briefings, rule events, and time anchors. Compare intent and actual execution; do not add a different tactic on behalf of the model. Limit calls, corrections, and retries as per the user's authorization; resume from breakpoints without silently swapping models or restarting.
4. Supervise before editing. Save the initial setup and full-map and restricted snapshots after each complete turn settlement, showing the current game/turn, key changes, and an index that can be opened. Pause for the next turn when the user requests "wait for my confirmation before continuing"; providing a snapshot per turn does not imply waiting for approval each time.
5. Extract facts before writing the script. Complete the agreed batch, retain all battle reports and statistics, then select games based on user preferences. Torpedo hit events, actual damage, sinkings, friendly fire, and scores are tallied separately. Explain selections using the model's own words or its public briefings; do not fabricate words or expose private reasoning.
6. Embed mechanics into events. Present distance, visibility, the time of old intelligence, and legal actions at the moment the audience needs to judge. For key choices, first show the restricted view of that seat, read the complete question aloud and pause for three seconds, then reveal the seat's actual choice and public reasoning, followed by cutting to the director's view to see the consequences.
7. Deliver a playable product. Verify files, images, subtitles, and audio; provide the completed MP4, subtitles, editable source, and cover. Briefly note missing items and acceptance boundaries in the delivery record; do not insert them into the narration.

Attached Offline Tools

Use only the Python standard library; do not request models, start competitions, or read credentials:

python scripts/episode_tool.py init /path/to/episode --episode-id episode-01 --games 5
python scripts/episode_tool.py ref /path/to/episode frames/game-1-t00-director.png
python scripts/episode_tool.py validate /path/to/episode --stage capture
python scripts/episode_tool.py supervision /path/to/episode
python scripts/episode_tool.py validate /path/to/episode --stage edit
python scripts/episode_tool.py validate /path/to/episode --stage delivery

First, map real engine output to the contract; do not use templates to generate fictional battle reports. Tools check file integrity, snapshot coverage, timeline, and three-second reveal relationships; they do not prove correct double-blind filtering, rule legality, or actual model calls — these must be verified at the engine/adapter layer. The supervision page contains the director's omniscient view, which is only for production supervision and must not be given to the command model.