rts-ai-doctrine

RTS AI Doctrine

Build RTS AI that feels deliberate rather than omniscient, reactive, or randomly scripted.

This skill focuses on layered AI for browser RTS games, especially single-page HTML/CSS/JS implementations where the same runtime owns simulation, economy, scouting, memory, and tactical orders.

When to Use

• Creating AI factions for an RTS
• Designing economy, production, and expansion logic
• Adding scouting, recon, and memory-based target selection
• Assigning tactical roles to scouts, tanks, support units, and aircraft
• Building phase-driven AI behavior such as normal play, siege, or resource-race modes
• Making AI use the same unit command systems as the player instead of cheating

Avoid Using This Skill For

• Server-authoritative PvP matchmaking AI
• ML-based opponent training systems
• Deterministic puzzle scripting without economy or tactical layers
• Simple wave spawners that do not make strategic decisions

Core Principles

• Separate economy, strategy, and tactics into distinct layers
• Let AI reason from memory and visibility, not perfect information
• Score objectives instead of hardcoding a single script
• Reuse the same order issuance helpers that player units use
• Prefer plausible, explainable behavior over maximum efficiency

Layered RTS AI Model

1. Economy Layer

This layer decides:

• what to build next
• what to produce next
• when to expand
• when to research
• when to stabilize income versus when to push military

Good economy AI usually depends on summary counts rather than raw scans each time.

Track and reason about:

• current resource stockpile and storage capacity
• counts by unit and building type
• unlocked tech and prerequisites
• shortages such as no refinery, no scouts, no engineers, or no anti-air

2. Strategic Phase Layer

Add explicit high-level phases, for example:

• normal
• defensive-recovery
• resource-race
• siege
• expansion

Phases should be triggered by measurable state:

• remaining map resources
• current income pressure
• base losses
• known enemy tech or threat concentration
• whether map control is slipping

This makes AI easier to tune and debug than one giant decision tree.

3. Tactical Assignment Layer

This layer assigns unit roles based on the current strategic phase and visible opportunities.

Examples:

• scouts to recon unexplored or stale areas
• engineers to structure placement or repair
• support vehicles to damaged allied clusters
• strike aircraft to visible high-value targets
• loiter aircraft to contested objectives or threat corridors

Keep tactical assignment separate from economy production logic.

4. Unit Execution Layer

Once a role is chosen, use the same order model as the rest of the game:

• move
• attack-move
• guard
• strike
• capture
• service

Do not build a second hidden AI-only movement system unless absolutely necessary.

AI Memory and Reconnaissance

Memory, Not Omniscience

AI should act on what it has seen or inferred.

A useful memory record stores:

• entity id
• kind and type
• last known position
• last seen time
• health if known
• any strategic tags such as “high-value” or “resource owner”

Reconciliation Rules

When the AI regains vision:

• update matching memories
• delete disproven memories
• downgrade confidence in stale information

This creates meaningful recon and prevents AI from feeling unfair.

Staleness and Decay

Do not let old intel drive behavior indefinitely.

Useful patterns:

• objective score decays with time since last seen
• stale threat estimates lose weight gradually
• recon tasks become higher priority when important memories are old

Objective Scoring

Use scoring rather than fixed if/else scripts for:

• expansion locations
• resource capture opportunities
• harassment routes
• defensive reactions
• air patrol points
• high-value strike targets

Common scoring inputs:

• distance from friendly hubs
• threat nearby
• value of the structure or unit
• whether the target is visible now
• how recently it was seen
• current strategic phase

Add hysteresis when helpful so the AI does not thrash between nearly equal objectives.

Production and Tech Heuristics

Use summary counts and capability gaps to drive production.

Examples:

• if no engineer exists, produce one before greedier tech
• if income is low and storage is blocked, prioritize economy infrastructure
• if airfields exist but no fighter is available, queue aircraft production
• if support ratio is too low for the combat force, add support units
• if the enemy fielded air threats, prioritize anti-air capable assets or scouts

Prefer explicit heuristics to invisible “weights” unless the weights are exposed and tested.

Air Doctrine

Aircraft need doctrine, not raw movement.

Useful doctrines include:

• direct strike on a visible high-value target
• loiter guard over a resource, objective, or remembered threat point
• return to service and relaunch to a queued mission

Recommended behavior split:

• one aircraft may commit to a visible direct strike target
• additional aircraft may loiter over the strategic objective or approach corridor
• AI should only assign direct strike when the target is current and meaningful

Scouts, Recon, and Resource Play

Scouts should not merely idle near base.

Good scout workflows:

1. unexplored grid sweep
2. stale-memory refresh
3. resource recon
4. harassment or capture when conditions are favorable

In resource-race or expansion phases, scout priorities should lean toward:

• resource node ownership
• capture feasibility
• local threat around neutral objectives

Tactical Safety Rules

Add explicit rules for fragile units.

Examples:

• engineers retreat from active threats
• ambulances avoid entering effective enemy range unless protected
• recon drones avoid redundant coverage if another unit already sees the area
• aircraft avoid landing or retasking into impossible service states

These small rules often matter more than broad AI strategy.

AI Workflow for New Features

When adding a new AI-capable system, follow this procedure:

1. Add the new unit or structure to the asset count summary.
2. Decide which strategic phase values it.
3. Decide what objective or role it serves.
4. Add a tactical assignment helper.
5. Reuse player-facing order helpers where possible.
6. Add safety and stale-intel rules.
7. Playtest with constrained scenarios, not only full matches.

Debugging Checklist

• Verify the AI can explain its current phase in terms of measurable state
• Verify it is not using hidden information the player could not know
• Verify objective scores change in understandable ways when the map changes
• Verify units are not being reassigned every tick without hysteresis
• Verify aircraft, scouts, and support units receive role-appropriate orders
• Verify stale memories are refreshed or abandoned over time
• Verify production heuristics do not deadlock on missing prerequisites

Example Prompts

• /rts-ai-doctrine Add layered AI phases for economy, recon, siege, and resource-race behavior.
• /rts-ai-doctrine Design memory-driven scouting and attack targeting for a browser RTS opponent.
• /rts-ai-doctrine Make AI fighters choose between direct strike and loiter patrol based on visible objectives.
• /rts-ai-doctrine Review my RTS AI for cheating behavior, brittle heuristics, and missing hysteresis.

Good Companion Skills

• rts-games for overall architecture and command semantics
• rts-pathfinding for movement, occupancy, and traversal logic
• rts-rendering-persistence for minimaps, fog, and save/restore systems