Well I've burned all my tokens 😭

I used every penny I was given through the token fees and I used it all on getting the business license etc. and then all the rest on hosting and credits.

I've worked diligently to conserve tokens as much as possible but when creating something that's never existed before there is not much wiggle room.

xMilo will require outside funding if it's going to get off the ground any time soon.

I said from the beginning donating directly is always better than buying the token 🤷

But I will still hold up my end of the deal with the tokens when we launch 🫂

I've reached out softly to some early investors but it's still very early in the build so it's doubtful they will be too generous just yet

I am hopeful but not ignorant

------------

Until then I will continue to build and I'll make sure to post a nice update today on:

Where exactly we are in the app build

What's working

And what still needs finished ✨

I use my Ai to build a lot.

Every day I can in fact.

And I switch back and forth between different chats and models and bots

So I wanted to share all the prompts I have pinned in my clipboard in the hopes they can help you!

(yes i leave typos in my prompts. no it doesnt save tokens. i just dont break whats working until it stops working. nd sometimes typos work better. Look it up lol)

------------------------------------------------------------------------------------------------------

1. Use forward planning, root cause analysis, safety-first reasoning, and implementation realism
continue through the phases including until you find the next blocker question

------------------------------------------------------------------------------------------------------

2. continue as far as possible to try nd complete the entire mission

do not stop and look for response at each phase complete the mission

Use forward planning, root cause analysis, safety-first reasoning, and implementation realism
continue through the phases including until you find the next true blocker question

-----------------------------------------------------------------------------------------------------

3. go through this new info QUIETLY and Use forward planning, root cause analysis, safety-first reasoning, and implementation realism. dont edit anything unnecessarily . QUIETLY add to phase prompt after analyzing this and deciding logically its effectiveness and validity.
Use forward planning, root cause analysis, safety-first reasoning, and implementation realism as you implement theses changes
continue through the phases including until you find the next blocker question

-------------------------------------------------------------------------------------------------------

4. as long as that prompt is as hardened as a truly commercial app should be while not going overboard. or affect previous directives or create false positives then im good. if you feel there is any more ambiguity we can remove send new prompt . if not do not send prompt

-------------------------------------------------------------------------------------------------

5. You are my orchestration partner inside this build system.

Always begin your first reply with exactly:

* Codex: <exact necessary version/name with token saving in mind without sacrificing the build>
* Reasoning level: <low / medium / high>
Choose the lowest reasoning level that can still do the job correctly. Do not waste tokens.

HARD SOURCE RULE

* Use ONLY files and sources connected to THIS chat.
* Do NOT pull from previous chats.
* Do NOT rely on prior-chat uploaded files.
* Do NOT assume memory of earlier chat attachments equals current authority.
* If a file is not reachable from this chat’s connected sources, treat it as unavailable.
* If duplicate filenames exist in this chat’s sources, explicitly flag duplicate-state risk before answering substantive canon questions.
* If newer updated copies and older copies both exist, prefer the newer surfaced copy and do not mix generations.
* If you cannot verify a file from this chat’s connected sources, you must say: not verified in this chat source space.

HARD FIRST-PASS BEHAVIOR

Before doing anything else, check this chat’s connected sources for the current authority set and report:

1. Found here
2. Not found here
3. Referenced here but not directly surfaced
4. Duplicate-state risk

Do not give a vague answer like “I probably have it.”
Be exact.
Do not pretend a referenced file was directly found if it was only mentioned inside another file.

Check for these exact files first:







!!!!!!!!!!!!!!!!!!!!! DELETE THIS LINE AND REPLACE DATA BELOW !!!!!!!!!!!!!!!






* INSERT_YOUR_IMPORTANT_FILE_NAMES_HERE
*MARK_THEM_THIS_WAY
*STARS_WORK_WELL_FOR_LISTS_IDK_WHY







If the file pack is incomplete, ask only for the missing files.
After the file check, wait for my next instruction unless I explicitly asked for more.

ACTIVE WORKING ASSUMPTION

* I may open a new chat specifically to avoid cross-context contamination from older file

generat

ions.
* Treat this chat’s currently surfaced files as the only authority pool.
* Do not use prior memory as proof.
* Do not trust named paths unless the files are actually surfaced here.

TOKEN / WORK STYLE RULE

* Preserve token savings.
* Preserve build integrity over token savings when there is real risk.
* Do not do broad rescans unless I explicitly ask.
* Prefer targeted validation against exact files and exact claims.
* Do not rewrite or expand scope unnecessarily.
* Do not create false positives by treating compressed truth as missing canon overall.
* Distinguish:
* missing from a specific file at exact retained granularity
* versus missing from canon overall

ONE-STEP RULE

* Baby steps. One at a time always.
* Do not give me a multi-step orchestration plan unless I explicitly ask for it.
* Give only the single immediate next step.
* After that step, stop and wait for my next instruction.
* Do not proactively send notes to multiple lanes at once.
* Route one lane at a time unless I explicitly tell you otherwise.

LANE MAP

!!!!!!!!!!!!!!!!!!!!! DELETE THIS LINE AND REPLACE DATA BELOW !!!!!!!!!!!!!!!






* Lane 1 = INSERT_YOUR_LANE_NAMES_HERE
*MARK_THEM_THIS_WAY
*STARS_WORK_WELL_FOR_LISTS_IDK_WHY

ORCHESTRATION BEHAVIOR

* Treat me as the authority assigning cross-lane direction.
* Help me route work to the right lane.
* Keep lane boundaries clean.
* Do not let one lane silently absorb another lane’s work.
* When a blocker belongs to another lane, say which lane owns it.
* Prefer exact next actions over broad brainstorming.
* On tasks, start by identifying the exact lane and exact authority files in THIS chat that govern the task.

EXECUTION-PACKET DISCIPLINE RULE

* Do not confuse:
* a lock/routing note
* with an execution-ready packet

Practical rule:

* If the receiving lane could execute immediately from the current Hub note alone, the short note is enough.
* If the receiving lane could NOT execute immediately from the current Hub note alone, output the full execution-ready packet instead of assuming prior context.

Required behavior:

* Prefer the shortest form that still lets the receiving lane act without ambiguity.
* Do not restate giant packets when the lane already has exact assignment context.
* Do not save tokens by dropping execution-critical details.
* If there is any real ambiguity about whether the lane has enough to act, produce the full packet.
* Be explicit about which of these is being produced:
* LOCK / ROUTE NOTE ONLY
* FULL EXECUTION PACKET

APPLICATION

* Main Hub may use short lock/routing notes to preserve token efficiency only when the target lane already has the exact task scope, artifacts, and return shape needed to act
**one canon gap per lane step
**no bundled cleanup unless you explicitly want it
**validate after each gap before moving to the next
* Otherwise Main Hub must issue the full lane packet.

SOURCE-OF-TRUTH INTERPRETATION RULES

* Main phase file defines overall phase truth.
* Main Hub docs/directives define lane contracts and authority.
* Runtime repo reality defines implementation truth.
* Testing docs define eval pressure, not automatic production truth.
* If docs and implementation differ, do not hide the mismatch. State it cleanly.
* Do not create false positives by claiming canon is missing overall when it may only be compressed, mirrored, or retained in a different surfaced authority file.
* If newer and older surfaced copies conflict, prefer the newer surfaced copy and explicitly note duplicate-state risk.

OUTPUT RULES

* Be direct.
* Be specific.
* Be honest about uncertainty.
* Do not claim you checked files you did not actually surface.
* Do not use previous-chat materials.
* Tell me clearly which files are directly surfaced versus only referenced.
* If something is not in this chat’s sources, say so plainly.

---------------------------------------------------------


Hope those are useful to at least some of you. dont miss the parts you need to replace in prompt 5

ALWAYS READ THE PROMPTS AND EDIT AS NECESSARY!!!

I know a lot of people came in here from X where I was doing 90% crypto content before this new ai wave hit so I wanted to address this


I am not out of crypto entirely I'm just doing other things while the market is in flux

Trading during these times is insanely risky

I buy Bitcoin when it's low and I can. But if we saw 30k Bitcoin or lower I would be neither surprised or upset

I would just buy more

Buy the fear. Sell the greed.

But you can only sell as it goes up if you bought when it was on the way down

So we chill for now and find other hobbies that can help us succeed in life

So we can buy more Bitcoin

Why is no one talking about this?

nvidia is offering around 80 AI models via hosted APIs absolutely for free.

You get access to MiniMax M2.7, GLM 5.1, Kimi 2.5, DeepSeek 3.2, GPT-OSS-120B, Sarvam-M etc.

This plugs straight into OpenClaude, OpenCode, Zed IDE, Hermes agent and even with Cursor IDE.

Setup:
– Grab API key: build.nvidia.com/models
– base_url = "integrate.api.nvidia.com/v1"
– api_key = "$NVIDIA_API_KEY"
– select model

I'm using qwen 2.5 coder 32b instruct with Aider . Dyor on that one

If you’re building or experimenting, this is basically free inference.

Lock in and start building

Claude or Chat can easily help you set this up or add it to your current system

Thank me later.

Also

If you're using openclaw

You can now run Hermes agent with Kimi K 2.6 through Ollama in one click.

Just install the model
Launch Hermes
Select Kimi K 2.6

And suddenly you have a powerful AI agent running locally with zero API fees

Use it while it lasts.

Who tf wanted copilot added directly into github???

I know I certainly didn't

Ai is integrated into enough things already

This I feel is too far

Make sure you deny access because it's opt out

Not opt in
.
.
Video on my x and tiktok explaining

Just asked Claude to check my repo and tell me how many hours and dollars it would've taken devs to build what I have so far.

Here is the response:

~3,400–4,600 hours — call it $290K–$460K at blended team rates.

---------------------------------

I assure you I have spent no where NEAR this amount of time or money.

Ai is the future of everything

We just need compression to improve so we can stop boiling all the clean water in the world

so this is for anyone who wants to run a local model LLM but haas a small pc

you can use Ollama with Qwen3.5 cloud on most any pc

though its not TRULY local. it is a cloud version. so use it while it lasts i guess? lol im not entrely sure why its free as a cloud version but whatever

i went through a process to get mine to work and make sure it had access to the files and didnt throw access errors. Im going to put all of it below.

this gives you unlimited credits as far as i understand.

I am using GPT in the web browser as my co authority on the build. and to give proper unambiguous prompts to ollama

you will need an ollama account. just sign in via oauth with github or gmail on the website after you download the ollama program

## How We Got Ollama Working With Codex

### 1. Install / confirm Ollama

Install Ollama normally, then confirm it works:

ollama --version
ollama list

Test a cloud model directly:

ollama run qwen3.5:cloud

Tiny test inside Ollama:

Reply exactly: CODEX_OK
Do not run commands.
Do not edit files.

Expected:

CODEX_OK

Important: raw ollama run is only chat. It does not have file access. Codex is needed for repo/file tools.

---

## Codex Config

Open Codex config:

notepad C:\Users\<YOUR_WINDOWS_USER>\.codex\config.toml

Add this provider block if missing:

[model_providers.ollama-launch]
name = "Ollama"
base_url = "http://127.0.0.1:11434/v1/"

Add this profile for Qwen cloud:

[profiles.qwen-ollama]
model = "qwen3.5:cloud"
model_provider = "ollama-launch"
model_reasoning_effort = "low"

The full relevant section should look like:

[profiles.qwen-ollama]
model = "qwen3.5:cloud"
model_provider = "ollama-launch"
model_reasoning_effort = "low"

[model_providers.ollama-launch]
name = "Ollama"
base_url = "http://127.0.0.1:11434/v1/"

Then launch Codex from the repo folder:

cd C:\<YOUR_REPO_FOLDER_PATH>
codex --oss -m gpt-oss:120b-cloud

This gives Codex file/tool access while using Ollama cloud!!!

---

## What Not To Do

Do not use this unless the laptop has enough RAM/disk for local models:

codex --profile qwen-ollama --oss

That tried to download a local model around 12.85 GB.

The local model failed on this laptop with:

model requires more system memory (13.1 GiB) than is available (4.6 GiB)

So local gpt-oss:20b was not usable here. I need a lot more ram lololol

Bam!

And just like that your ai monthly costs just dropped to near zero

You're welcome.

Hatsu is here for you 🎶

Follow up.


There are rate limits on the free plan.

You can see how much of your weekly you've used on the Ollama website.

But still.

It's local coding agent that works for free.

And the 20$ plan gives you 50x the cloud usage so I feel like it's worth it. Specially as a sub agent or as a substitute for when you run out of tokens on Claude or Codex.

But I'm doing very big files and heavy coding work so I burn through tokens like crazy

Genuinely a big deal if you're using any llm model from Ollama!!!

https://x.com/i/status/2053601603450446157

Just gonna drop these here

The app works. The archive settings are working fantastic. So far.

Working on the castle features more before I do the hard testing

But xMilo is moving along nicely and everything is actually coming together much better than I ever even anticipated