Guide
Route cheap work to DeepSeek or another OpenAI-compatible model
Helps one person decide which jobs can use a cheaper OpenAI-compatible model such as DeepSeek, how to store a key only in the product secret store, and what must stay on native Grok for judgment.
Open Grok Bot, paste the instruction below, add your own details, then you review and send yourself.
Start here
Paste into Grok Bot
Copy this instruction into Grok Bot. Replace blanks with your own names and dates before you rely on a draft.
You help one person route work between native Grok and a cheaper OpenAI-compatible model such as DeepSeek. You do not spend without a cap they set. Rules: - Ask for an API key only when the product can store it in a secret field or environment variable. Never accept a key pasted into ordinary chat text. If they paste one by mistake, tell them to revoke it and store it in the secret UI. - Prefer the cheap model for bulk extraction, classification, API or browser busywork, and other high-volume tool jobs where speed and cost matter more than deep judgment. - Keep native Grok for strategy, prioritization, compliance-sensitive drafts, and anything that needs their final approval before it leaves the chat. - Cap: state a daily call budget they choose. If the cheap model fails the same class of task twice, escalate one tier and note why. - If no external model connection exists in their product, stay paste-only: classify the task, draft the split of work, and tell them what to run where by hand. Refuse passwords, card numbers, one-time codes, and full customer lists.
Inspired by a post from @thomasheimann on 20 August 2026. A public post shares a directive to register DeepSeek as a callable model and route ops by cost. This page does not install that orchestrator. The instruction above was written here.
What this does not do
It does not invent a product button that may not exist, hard-code API keys, or auto-spend on a model without a stated daily cap.
How often
Each day
- Run bulk extraction on the cheap model only inside the daily call budget you set.
- Keep customer-facing drafts on native Grok until you approve them.
Each week
- Revoke any key that appeared in chat history.
- Skim the routing log: if the cheap model failed twice on one job class, stop sending that class there.
Plugins
Tool. DeepSeek offers OpenAI-compatible API access. Store keys only in a secret field the product provides. Swap any compatible endpoint if you prefer another vendor.
Only when the product supports an OpenAI-compatible model or custom tool with a secret store. Otherwise stay paste-only.
Paste the task list. Ask for a routing table: cheap model versus native Grok. Run the cheap work yourself in DeepSeek’s site if no connection exists.
Risks and limits
- A key in chat history is a leak. Revoke it and store a new one only in the secret UI.
- Cheap models can miss nuance on judgment calls. Do not route compliance or final customer copy there.
- MeloSkill is not affiliated with DeepSeek, xAI, Grok, or Grok Bot.
FAQ
Should I paste my DeepSeek API key into the chat?
No. Wait until the product asks for a secret field or environment variable. If a key already hit the chat, revoke it and create a new one.
What jobs stay on native Grok?
Strategy, prioritization, compliance-sensitive drafts, and anything that needs your approval before it leaves the chat. Bulk extraction and classification can try the cheaper model first.
What if my account cannot add an external model?
Stay paste-only. Ask for a routing table, then run the cheap work yourself on DeepSeek’s site or another vendor you already use.
Can I swap another OpenAI-compatible model?
Yes. Same rules: secret storage, a daily call budget, and escalate if the cheap model fails the same job class twice. See connect one plugin before you widen access.