Data.
Your offers, customers, examples and evidence. Give the work a reliable starting point.
Give it context →T H E · H O L Y · T R I N I T Y
Data + AI + Automation.
Your AI operations.
Sorted.
Your workspace. Your specialist team.
An assistant built around the way you work.
A L R E A D Y · S E T · U P ?
Build an actually useful assistant.
Turn real work, your voice and your standards into instructions your AI can follow.
Paste into your connected project after Step 2.
S T A R T · H E R E · M A C
One command.
The whole setup.
Creates AI Operations, adds the assistant playbook and six role templates, installs MangoMagic 7.1 through Ollama, then registers it with ChatGPT.
PASTE INTO TERMINAL
bash -c 'f=$(mktemp) || exit; trap "rm -f \"$f\"" EXIT; curl -fsSL https://raw.githubusercontent.com/mango-magic/mangomagic/main/setup-operations.sh -o "$f" && bash "$f" --with-mangomagic'Automatically quits and reopens ChatGPT. Allow any required account sign-in, then follow Steps 2 and 3 below to connect your folder and personalise your assistant.
Want just the folders and assistant kit?
Use this if you already have a model. You can add MangoMagic later in Step 6.
WORKSPACE ONLY
bash -c 'f=$(mktemp) || exit; trap "rm -f \"$f\"" EXIT; curl -fsSL https://raw.githubusercontent.com/mango-magic/mangomagic/main/setup-operations.sh -o "$f" && bash "$f"'T H E · H O L Y · T R I N I T Y
Data + AI + Automation.
One connected way to work.
AI.
A specialist with a clear job, the right sources and a definition of done.
Give it ownership →Automation.
Test the job once. Put the useful parts on repeat. Hear about the changes that matter.
Give it a routine →F R O M · F O L D E R · T O · F I N I S H E D
Your next steps, in order.
If the everything command completed successfully, Steps 1 and 6 are done. Continue at Step 2, then build your assistant in Step 3. If setup stopped, resolve the error shown and rerun it.
S T E P · 0 1
Step 1. Build the workspace.
Open Terminal on your Mac. Paste this once. The script creates Documents/AI Operations with six role folders, project templates and the complete Build My Assistant playbook.
PASTE INTO TERMINAL · MAC
bash -c 'f=$(mktemp) || exit; trap "rm -f \"$f\"" EXIT; curl -fsSL https://raw.githubusercontent.com/mango-magic/mangomagic/main/setup-operations.sh -o "$f" && bash "$f"'No Git or Python needed. Existing files are preserved. It opens your folder in Finder; it does not launch agents, schedule jobs or change your global settings. Read the script ↗
Prefer a download, or a different folder?
Download the starter ZIP and unzip it. To choose a destination with the script:
TERMINAL
bash -c 'f=$(mktemp) || exit; trap "rm -f \"$f\"" EXIT; curl -fsSL https://raw.githubusercontent.com/mango-magic/mangomagic/main/setup-operations.sh -o "$f" && bash "$f" --destination "$HOME/Desktop/AI Operations"'The command is for macOS. On other systems, use the ZIP and a client that supports the workflow.
S T E P · 0 2
Step 2. Connect ChatGPT.
- In the desktop app, create a local project called AI Operations.
- Open the project menu: Edit project → Add folder. Choose your new AI Operations folder and make it primary.
- Start a new chat in that project. Continue to Step 3 and paste the assistant prompt.
The first check is real file access. If the assistant cannot read your mission and business brief, setup is not connected yet. Official project instructions ↗
Using ChatGPT on the web?
Create a ChatGPT project, upload the relevant starter Markdown files and business sources, and paste the shared AGENTS.md rules into project instructions. Keep uploads current. A web project does not automatically read your Mac's folders, run the Mac installer or inherit its Ollama model. Local custom agent definitions apply to supported local Codex clients.
Use the prompts with the tools your account actually has. If file writing or delegation is unavailable, the assistant can return file contents and work through roles sequentially.
S T E P · 0 3
Step 3. Build an actually useful assistant.
Your assistant should know what you do, how you like to work, what good looks like and when it can act. Our Team AI Operations method turns those answers into a personal operating brief.
- Understand your actual week. Use recent tasks, real examples and the work that never makes the calendar.
- Learn your working preferences. Answer first or detail first, how to deliver bad news, and examples of your voice. No personality labels in its replies.
- Make the work specific. Give each task a trigger, inputs, steps, output, quality check and permission level.
- Prove it on real work. Produce one useful deliverable, then choose three tasks for the first week. Record gaps and check the results.
PASTE INTO YOUR CONNECTED PROJECT CHAT
Build an actually useful assistant for me in this AI Operations workspace. Read START_HERE.md, AGENTS.md and 00_Command_Centre/BUILD_MY_ASSISTANT.md. First verify that you can read the files; if access is missing, say exactly what is missing. Use the context and permissions I have already given you. Ask only for material gaps, up to five concise questions together, and follow my preferred pace. Learn from real tasks and examples of my writing. Adapt your communication to my preferences without mentioning personality types. Follow the playbook to build my profile, specific task register and reusable assistant instructions. Mark facts, assumptions and gaps honestly. Preserve existing work, update project_tasks.json, and produce one useful first deliverable with checks. Prepare three practical tasks for week one and the rollout audit. Account connections, external actions and schedules need their actual tools and my authorisation; never report them active just because files exist.
Read the full assistant guide ↗ · Download the playbook ↓ · Get the audit checklist ↓
Already installed an earlier kit? Rerun the workspace command to add the new assistant files. Your existing files stay intact. On the web, upload the assistant Markdown files to your project first.
Know where everything lives.
- 00_Command_Centre
- Your assistant playbook and profile, business brief, task brief and handoff template.
- 01_Data
- Source documents, sanitised conversations and a source register.
- 02_Agents
- One folder and instruction file for each role.
- 03_Projects
- One folder per outcome. Drafts, task evidence and checks.
- 04_Automations
- Tested routines and records of actual schedules.
- 05_Deliverables
- Reviewed work you can find, use and share.
At the root: AGENTS.md sets working rules, CURRENT_MISSION.md sets the priority, and project_tasks.json records actual progress.
S T E P · 0 4
Step 4. Give everyone a job.
Use one lead to coordinate. Bring in specialists when the task benefits. Start with one useful outcome, not six busy agents.
Chief of Staff.
Turns priorities into owned tasks and brings the finished work together.
Sales.
Works from buyer conversations, approved offers and real objections.
Research.
Finds sources, checks claims and makes uncertainty visible.
Content.
Turns approved ideas into clear drafts in your voice.
Automation.
Makes the repeatable parts reliable, with a test and a stop condition.
Quality.
Checks whether the work meets the brief and the evidence supports it.
Each role gets an AGENTS.md and a working folder. The kit also includes project-scoped .codex/agents definitions for supported local clients. The onboarding prompt checks what is available. Folders hold context; delegation starts workers.
Already have agents? Use ROLE_TEMPLATE.md and add a matching folder to the roster. Keep the names you already use.
Official subagent guidance ↗ · Each worker consumes model usage. More workers are useful only when their jobs can run independently.
S T E P · 0 5
Step 5. Put good work on repeat.
Run a task manually. Check the output. Fix the instructions. Then schedule it. Start with a daily priorities brief, a weekly team review or a delivery check.
- Use the automation prompt below to define inputs, cadence, timezone, output and failure handling.
- Test the prompt in a normal project chat.
- Create the task in the app and verify it under Scheduled, including its timezone and enabled status.
- Review the first run. Keep the useful notifications; cut the noise.
For local-file schedules, keep the computer on, the app running and the project available. A Markdown file is a reusable brief, not a running schedule. Official scheduling guide ↗
Copy the automation prompt ↓S T E P · 0 6 · O P T I O N A L
Step 6. Add MangoMagic 7.1.
MangoMagic 7.1 by ManyMangoes. The business AI system for the modern worker. Bring files, screenshots and real work. Get concise answers, checked outputs and useful delegation through your app’s available tools. This step pulls the model, checks it, registers it with ChatGPT and gracefully restarts the app. Skip it if you only want the workspace.
TERMINAL · RESTARTS CHATGPT
bash -c 'f=$(mktemp) || exit; trap "rm -f \"$f\"" EXIT; curl -fsSL https://raw.githubusercontent.com/mango-magic/mangomagic/main/install.sh -o "$f" && bash "$f"'After the restart, choose MangoMagic 7.1. Images are supported. Light = Low/Light, Mango = High, Super Mango = Max. ChatGPT owns the slider labels.
All three use the same model and token rates. Extra reasoning can use more time and tokens; it does not guarantee a better answer. Ollama cloud usage is separate from your ChatGPT subscription. This is instruction customisation, not weight fine-tuning.
C O P Y · P A S T E · P R O G R E S S
Give it a job. Get an output.
These go into a project chat, not Terminal. Replace the bracketed details.
T E A M
Brief the specialists.
One outcome. Clear ownership. One integrated result.
Read prompt
Read AGENTS.md, CURRENT_MISSION.md and project_tasks.json. Deliver [specific outcome] in 03_Projects/[project]. Use [source files] and meet [acceptance criteria] by [deadline]. If supported, delegate independent parts to up to three relevant specialists. Give each a role file to read, exact inputs, a separate output path, file scope and a definition of done. Use fewer agents when practical. If delegation is unavailable, work sequentially and say so. Integrate the work, verify it against the evidence, update the tracker and return the deliverable with any remaining decisions. Continue authorised work without unnecessary check-ins.
S A L E S
Turn conversations into a playbook.
Find what moves buyers forward, and what does not.
Read prompt
Read the approved offer and sanitised inbox conversations in 01_Data. Build a sales playbook for [buyer segment]. Find buying triggers, objections, useful questions and next steps. Cite anonymised source IDs and show sample size, counterexamples and whether a meeting was actually confirmed. Separate observed patterns from your suggestions. Do not invent conversion rates or personalisation. Produce three short draft response sequences, verify every business claim, and save the playbook and evidence in 03_Projects/[project]. Update project_tasks.json. Do not send messages.
R E S E A R C H
Make a decision with evidence.
Useful options, checked claims and a recommendation.
Read prompt
Research [decision] for [audience] using [approved documents and research scope]. Verify time-sensitive facts with primary sources when research tools are available. Record source dates and links. Separate facts, inference and unknowns; include conflicting evidence and what could change the answer. Save a one-page decision brief with options, recommendation, key risks and next action in 03_Projects/[project]. Check every material claim against its source. Update project_tasks.json and state any sources you could not access.
D E L I V E R Y
Keep going. Check the work.
Resume from the files and finish what is authorised.
Read prompt
Resume the current mission from AGENTS.md, CURRENT_MISSION.md, project_tasks.json and the latest handoffs. Inspect the actual outputs before trusting recorded status. Complete the highest-priority unblocked work within the authorised scope. Verify acceptance criteria and record evidence paths, failures and next actions. Never inflate completion percentages. Prepare the finished deliverable for owner review only after all required work and checks are done. If a decision or missing access blocks progress, state exactly what is needed and continue independent work. Return the result, its verification and the next decision.
A U T O M A T I O N
Build a routine worth keeping.
Prepare it, test it, then verify the actual schedule.
Read prompt
Prepare a recurring [job] for this project. Cadence: [schedule]. Timezone: [Area/City]. Inputs: [specific accessible files or connected tools]. Output: [file and success check]. Read 04_Automations/AUTOMATION_BRIEF.md and fill a copy for this routine. Define duplicate prevention, retry limit, stop condition and usage budget. Notify only on a meaningful change, actionable failure or required decision. Run one bounded manual test and inspect its output. Keep this a draft until I ask to schedule it; a saved file is not a scheduled task. Then verify the saved app schedule, timezone, enabled status and input access before reporting it active. Do not send messages, publish or spend as part of this routine.
Q U A L I T Y
Give it a proper final check.
Find material problems before they reach a customer.
Read prompt
Review [deliverable path] against [task brief] and its cited sources. Check factual claims, calculations, links, missing requirements and whether the output is useful to its intended reader. Distinguish checks actually performed from checks still needed. Return pass, needs changes or blocked, with evidence and specific corrections. Fix routine issues within the authorised scope, recheck them and update project_tasks.json. Keep owner approval separate from technical completion. If this reviews your own work, disclose that the review is not independent.
K E E P · T H E · P L A Y B O O K
Files that do the explaining.
Read them, adapt them, or give them to your project chat. The starter contains generic templates, not anyone's private business data.
Download starter kit ↓W H E N · S O M E T H I N G · D O E S N ' T · C L I C K
A few useful answers.
Does the everything command finish my assistant too?
It installs the files and model integration. Step 2 gives the project access to your folder. Step 3 uses your real work and preferences to build your assistant. An installer cannot answer those questions for you. The assistant prompt handles the setup conversation and checks its first result.
It says setup finished. Why does ChatGPT not know my files?
The script finished creating the workspace. Attach it as the primary folder of a local project and start a new chat there. Ask ChatGPT to name and summarise the mission file. A file existing on disk does not prove the app can read it.
Do I need MangoMagic for this system?
No. The workspace and prompts can be used with the model available in your app. MangoMagic 7.1 is ManyMangoes' business AI system for the modern worker, with images and adjustable thinking. Model availability, tools and scheduling depend on your client and account.
Will this use my existing agents?
It preserves existing files and global settings. Map your agents to folders in AGENT_ROSTER.md using ROLE_TEMPLATE.md. The six included project definitions are suggestions, not an import of your existing account configuration. Ask the app to verify the roles it can actually delegate to.
Is this training the model on my business?
It supplies context and instructions. It does not change model weights or make a public model know your private files. Only provide source material you want the selected model and connected tools to process.
What should I automate first?
A short internal priorities brief or a weekly delivery review. Give it a narrow input, a useful output and a check you can repeat. Test it before adding a schedule. Keep customer messaging and publishing within your explicit authorisation.