AI Scenario Help

[Lobster]

Just for the heck of it I decided to ask MS CoPilot to make a scenario of the 1939 invasion of Poland using The Operational Art of War IV. I was gobsmacked. The level of detail was remarkable. First it was just a general over view but I kept asking for more detail and it just dug deeper and deeper. You can make an entire scenario with accurate TOE, events, OOB, deployment plan...everything. What it can't do is make the map but it will make you a bitmap image.

[Cpl GAC]

Yes - https://forums.matrixgames.com/viewtopi ... 3#p5261043 .

It gets down to amazing details to consider - like where stragglers from the Barbarossa border battles went, and therefore things like the 22nd Army at the Dvina/Dnepr should have more proficiency than the 20th Army because they got officer cadre from the 3rd Army;

" TOAW IV TRANSLATION (army-specific)
If you’re modeling unit survival:
- 3rd Army units:
- Reduce strength to 30–35%, keep HQs alive
- 4th Army units:
- Remove most formations; convert survivors to replacements
- 10th Army units:
- Disband entirely; feed manpower pool only
- 13th Army units:
- Reconstitute at 25–30%, high disruption "

Edit - and things like the Western Front in July should only get 1,200-1,500 Artillery pieces from the border stragglers and those were almost all; 50mm mortars, 45mm AT and 76mm Art - the easiest to move

[cathar1244]

and things like the Western Front in July should only get 1,200-1,500 Artillery pieces from the border stragglers and those were almost all; 50mm mortars, 45mm AT and 76mm Art - the easiest to move

Didn't mention 37mm "spade mortars" ?

Don't know. The logic of the suggestions sounds good, but I wonder how it actually worked out.

[Cpl GAC]

Well, AI said that if Lobster had actually included 37mm mortars in Campaign Barbarossa, it would be 338 of the 1,500 pieces. I was surprised that it referred to him by his code name and not his real name, nor the name Bob probably calls him under his breath. Then we went out for a beer.

AI can do some heavy lifting up front, but you have to know when it's wrong and where to fine-tune. Tangentially, it's not a very good writer, but I still pass drafts through it to see if its suggestions improve my writing.

[Simon Edmonds]

Yep. It blows my mind too. I have been doing some research on Danish towns using Google Gemini. This morning it comes up "this town is similar to one you asked about yesterday....." How many people would be using Google at any one time in the world. To remember a question that I asked yesterday (I was not logged into Google at the time) and link it to one I asked today (still not logged on) is just beyond belief. I would be interested to see the screen shot of the bitmap your AI came up with, Lobster.

[Curtis Lemay]

First, excellent use of the word "Gobsmacked"! You don't generally see that level of vocabulary here on the Matrix Games board.

However, I wonder if a better use of this AI resource would be actually playing the game in lieu of the PO?

[Lobster]

As in, "Would you like to play a game?"

[Lobster]

First, excellent use of the word "Gobsmacked"! You don't generally see that level of vocabulary here on the Matrix Games board.

However, I wonder if a better use of this AI resource would be actually playing the game in lieu of the PO?

I asked if Copilot can play a TOAW IV scenario. The answer:

Short answer: no — Copilot can’t “play” a TOAW IV scenario.
But it can do a lot around the game that makes designing, analyzing, or planning scenarios dramatically easier.
Here’s the breakdown:
What Copilot cannot do
- It can’t run TOAW IV or interact with the executable.
- It can’t load or play a scenario file (.sce, .sal, etc.).
- It can’t take in-game actions like moving units, resolving combat, or issuing orders.
- It can’t read or parse a scenario file directly from your computer.
TOAW IV is a closed Windows application with no API, so there’s no way for me to “play” inside it.
What Copilot can do — and this is where it shines
Given your scenario-design focus, these are the things I can help with at a very high level:
1. Scenario design support
- Build historically accurate OOBs with TOAW-ready stats.
- Generate event scripts, triggers, reinforcement schedules.
- Suggest victory conditions, PO strategies, and balance adjustments.
- Help you structure the scenario narrative and briefing text.
2. Map creation workflow
- Help you size and scale your map.
- Produce reference-map workflows and step-by-step editor instructions.
- Advise on terrain density, road networks, rail lines, and river hexes.
3. Operational analysis
- Evaluate historical plans and propose alternative deployments.
- Analyze force ratios, movement rates, and likely operational outcomes.
- Help you tune proficiencies, readiness, and supply levels for realism.
4. Play-by-play reasoning
If you describe:
- the map,
- the units,
- the situation,
- and the options available,
I can walk through what an optimal move might be, almost like a wargaming assistant or staff officer.
But that requires you to tell me what’s on the screen — I can’t see or access the game.

If you want, we can even simulate a turn-by-turn “AI opponent” where you describe the board state and I respond with plausible German or Polish operational decisions based on doctrine and TOAW mechanics.
If that sounds useful, tell me what scenario state you want to explore next — deployment, Turn 1, or a specific sector of the front.

[rhinobones]

But that requires you to tell me what’s on the screen — I can’t see or access the game.

Does it get the screen information from simply reading the .gam file?

Rgards

[Lobster]

I don't think the AI can read any of the game files. Not tried to have it focus on an xml. Now that's an idea. Wonder how that would work out.

[cathar1244]

I guess the key to AI utility in these cases is the amount of information available online. Looking for the order of battle for obscure campaigns could yield meager results. IIRC, the orders of battle and other information for Barbarossa are online, at Wikipedia and other sites.

[Telumar]

First, excellent use of the word "Gobsmacked"! You don't generally see that level of vocabulary here on the Matrix Games board.

However, I wonder if a better use of this AI resource would be actually playing the game in lieu of the PO?

Can you program an API for an AI to access the game as it is now? Just theoretically

Related: Have you read what they did with the AI for "War in Spain"? They had a "true AI" learned/trained the game. See here: https://www.matrixgames.com/news/war-in ... ev-diary-3
Won't that be an approach for a possible future incarnation of TOAW? One would need to reprogram it from start and thus this would require a major effort from Matrix i assume. I cannot see that anywhere on the horizon though... :/

[Curtis Lemay]

First, excellent use of the word "Gobsmacked"! You don't generally see that level of vocabulary here on the Matrix Games board.

However, I wonder if a better use of this AI resource would be actually playing the game in lieu of the PO?

Can you program an API for an AI to access the game as it is now? Just theoretically

Related: Have you read what they did with the AI for "War in Spain"? They had a "true AI" learned/trained the game. See here: https://www.matrixgames.com/news/war-in ... ev-diary-3
Won't that be an approach for a possible future incarnation of TOAW? One would need to reprogram it from start and thus this would require a major effort from Matrix i assume. I cannot see that anywhere on the horizon though... :/

I have some ideas on improving the PO, but that's as far as I can take it. Waaaaay down the road, AI opponents may be available on line to just read the rules and play any game there is. I thought that this might have been the first inklings of that.

[Spale69]

For the purpose of my work and hobbies (both related to software development) I do follow this whole story about AI. My opinion is that these tools in their default, especially non-payed mode cannot be used as strategy/tactical game advisors really to the best of their capabilities. I did have payed ChatGPT access last year and have tested it against a scenario played solitare in Vassal ("BoB - Ghost Panzers") and another one for few rounds with TOAW 4 (Kasserine). It showed some use, especially as overall adviser after uploading screenshots, first of terrain-only, and then of units. However, I thought that it is not so fun as when I play those games solitaire for my own fun (which means, I like myself to analyze sitation and find a solution).

Now, since then A LOT has changed in the capabilities of these tools. Especially with the payed versions where you can upload and maintain a context of what you are working with. Imagine it as an online folder of infos that you input into the tool (manuals, descriptions, secondary clarifications in discussions with the bot after its preliminary analysis...anything you consider important even remotely). If you build your context with the lot of relevant information and in clear, simple explanatory style, the whole AI thing becomes quite powerful and useful. In ChatGPT, basically that sort of architecture is called "project"...sort of a workspace. And full access to its capabilities you get with Pro (20$ currently). But still...you need to read about how to "talk" to the bot in most cohesive and explanatory way (people call it nowadays "prompt building techniques"). And even when you build that, the maintenance, further evolution and building of such context - and its self-learning capabilities - depends on how much resources you pay to your AI "provider": the more bucks you invest - the bigger context and other capabilities you can get.

Just by using those entry-level simple discussions and questions - you won't get far, don't waste time. That is meant - my feeling - as a sofisticated freeware demo. Unless you are happy with some overall abstract advices and/or bunch of mistakes, hallucinations and mistakes once the scenario advances. But even on that level - you need to know how most effectively to "talk" to that bot, to upload images - in general - to plan your approach and before everything, experiment to get the feeling what would be the best way to engage with the tool.

Anyhow, since I do have payed access again, I do plan to do exactly these kind of things for some simple WDS Panzer Campaigns scenario and one simple TOAW. Currently, I am too busy with my work and my "major" game and modding and our community - so that is eating all my little free time. Once I do this above, or at least start after some initial planning, for sure I will report here.

But - let's allow AI to tell for itself what it thinks about these topics. In the following post, I will paste my little chat with it. Also, do notice the way I structure my topic, questions, insecurities and also mentioning already in the introduction existing ChatGPT tools ("projects" and "custom GPTs"). Also, the bulleted list of what I was quickly thinking I can supply/submit for making a bigger context. In that way, you actually "lead" that algorithm which people call "AI" in certain direction.

Cheers. Sorry for longer post - but soon or later, I do believe (and am convinced from what I see internally in my work for one software company) - these bots will be way better opponents than most of the people. And crazy good analysts. But only with the built context.

[Spale69]

You said:
I am considering an idea to use somehow ChatGPT Plus (payed version) as an AI opponent and strategy/tactics advisor in turn-based hexmap wargames like "The Operational Art of War IV" (TOAW 4). What do you think about this? Any suggestions? I was thinking if maybe ChatGPT project can be better suited for one game run. Or maybe custom GPTs, I am not sure. I can eventually do the following to supply some knowledge:

- PDF manual of the game
- Map of the scenario (screenshot image, as well as further screenshots as game progresses)
- Description of the scenario
- On the map there can be counters visible with attack and defense strength
- I can supply information about turns when reinforcements arrive

Do you have also ideas what else can I bring?

ChatGPT said:
That’s a very interesting idea—and TOAW IV is actually one of the better candidates for this kind of experiment because it’s turn-based, information-dense, and already abstracted into operational concepts that an LLM can reason about.

I’ll break this down into (1) feasibility & expectations, (2) best way to structure it in ChatGPT (Projects vs Custom GPTs), (3) what you should feed it, and (4) practical workflow suggestions, plus limitations to plan around.

1. Is this a good idea? (Short answer: yes, with the right expectations)
What ChatGPT can do well

Act as a staff officer / operations planner

Provide course-of-action (COA) analysis

Identify operational objectives, risks, and trade-offs

Evaluate force ratios, terrain, supply, timing

Help avoid tunnel vision and “human bias”

What it cannot reliably do

Optimize at the micro-tactical level (exact attack sequencing)

Track dozens of counters perfectly across many turns

Replace a true rules-engine AI (it won’t simulate combat outcomes)

Think of it as:

“A human general with perfect doctrinal knowledge but imperfect situational awareness unless you feed it clean data.”

Used this way, it can be very powerful.

2. Project vs Custom GPT: what fits best?
ChatGPT Projects → best choice for a single campaign
I strongly recommend Projects for your use case.

Why Projects fit TOAW campaigns well

Persistent memory across turns

You can upload:

Manuals

Scenario briefings

Maps

Turn logs

You can evolve a single “campaign brain” over time

Easier to iteratively refine prompts

Think of it as your AI general’s war diary.

Custom GPT → better for reusable doctrine
Custom GPTs are useful if you want:

A reusable TOAW doctrine assistant

A fixed personality (e.g. “Soviet operational doctrine 1943”)

Preloaded assumptions about hex warfare

Best approach

Use a Custom GPT for:

Doctrine

Rule interpretation

General TOAW advice

Use a Project for:

A specific scenario and playthrough

You can even combine both (consult doctrine GPT → feed conclusions into the Project).

3. What to feed it (you’re already thinking in the right direction)
Your current list is solid. I’ll refine and expand it.

A. Core inputs (essential)
You already listed most of these:

Game manual (PDF)
Focus the AI’s attention by:

Explicitly stating which optional rules are ON/OFF

Highlighting combat resolution, supply, interdiction, and ZOC rules

Scenario description

Victory conditions

Time limits

Reinforcement schedule

Political / withdrawal constraints

Map images

Initial map (high resolution)

Subsequent turn screenshots

Tip: annotate screenshots when possible (arrows, circles, notes).

B. What you should definitely add (high impact)
1. Structured force summaries (text, not just images)
Images alone are limiting.

Every turn (or every few turns), provide tables like this:

FRIENDLY FORCES (Turn 6)
- III Corps: 6 divisions, avg atk 12 / def 10, supply 80%
- Armored Group A: 2 tank divisions, atk 20, fuel constrained
- Artillery Reserve: 3 units, attached to HQ X

ENEMY FORCES (Observed)
- Estimated 8 divisions in Sector North
- Strong defensive line along river hexes 23–30
This massively improves reasoning quality.

2. Command constraints
Tell it:

Stacking limits

HQ command range

Supply source locations

Rail conversion status

Movement penalties

Operational AI lives on constraints.

3. Your intent
Very important.

Example:

“My strategic intent is to break through in the south and encircle the enemy by Turn 12, even at higher casualties.”

or

“I prefer conservative play and preserving elite units.”

This helps it tailor advice.

4. Fog of War assumptions
Be explicit:

What is known

What is estimated

What might be hidden

Example:

“Enemy reserve strength east of the river is unknown.”

C. Advanced / optional but powerful inputs
1. Loss tolerance
Give numerical or qualitative limits:

“Acceptable losses: up to 20% strength”

“Avoid losing armored divisions”

2. Rules interpretation clarifications
Whenever a rule question arises:

Ask ChatGPT to quote the manual section it is using
This keeps reasoning grounded.

3. Turn log
After each turn:

TURN 7 RESULTS
- Attacks succeeded in 3/5 hexes
- Lost 1 infantry division (retreated and shattered)
- Enemy counterattack failed in sector B
This allows cumulative learning.

4. Practical workflow (this matters a lot)
Here’s a realistic and efficient loop:

Pre-game
Upload:

Manual

Scenario briefing

Ask:

“Summarize key rules and scenario-specific constraints that will shape operational planning.”

Each turn cycle
Provide new screenshots

Provide structured text summary

Ask one of these:

“Evaluate current operational situation”

“Propose 2–3 COAs with risks”

“Identify enemy vulnerabilities”

“What would you do this turn and why?”

Then you execute the turn.

Periodic deep review (every 5–10 turns)
Ask:

“Are we drifting from our original strategy? What adjustments are needed?”

5. Using ChatGPT as an AI opponent (harder, but possible)
This is trickier but still doable.

Best method

Role-play:

“You are now the Axis commander. Given this map and intel, what is your likely plan?”

You can:

Ask it to predict enemy priorities

Use it to sanity-check your own assumptions

Don’t try to let it “run both sides” fully—confirmation bias will creep in.

6. Key limitations to be aware of
It won’t remember exact hex coordinates reliably over dozens of turns

Visual parsing of dense counters is imperfect

It may hallucinate rules if not grounded in the manual

It cannot truly simulate TOAW’s combat engine

Mitigation

Use text summaries

Frequently restate constraints

Ask it to reason step-by-step

7. Bottom line
This is:

Absolutely viable

Best framed as an AI general staff, not a micromanaging AI

Most effective using Projects

Dramatically improved by structured text data

If you want, next we can:

Design a standard “turn input template”

Create a doctrine prompt tailored to TOAW IV

Define a consistent AI personality (e.g., cautious Soviet, aggressive German)

This is a really cool idea—and you’re already thinking about it the right way.

[Simon Edmonds]

I don't think the AI can read any of the game files. Not tried to have it focus on an xml. Now that's an idea. Wonder how that would work out.

I asked Gemini about auto generating a hex map and besides giving me a REALY complicated answer it demonstrated a firm understanding of the XML map files and what each field did. What I need to do is to learn how to ask the right questions, the right way.

[Lobster]

I don't think the AI can read any of the game files. Not tried to have it focus on an xml. Now that's an idea. Wonder how that would work out.

I asked Gemini about auto generating a hex map and besides giving me a REALY complicated answer it demonstrated a firm understanding of the XML map files and what each field did. What I need to do is to learn how to ask the right questions, the right way.

Reminds me of the movie 'I Robot'. 'That, Detective, is the right question'.