Introduction
“I’m paying for Claude Pro, but I keep hitting the usage limit…”
“I want AI to write code for me, but I don’t know which tool to use…”
Do you have these kinds of concerns?
In this article, we’ll introduce a workflow that accelerates development by combining free AI agents while minimizing the consumption of tokens from paid tools like Claude and GitHub Copilot.
What You’ll Learn
- A workflow design that avoids wasting paid AI tokens
- How to effectively delegate tasks to free AI agents (Jules / Antigravity / Gemini CLI)
- The real characteristics, success rates, and limitations of each tool
Prerequisites: Tool List
1 As of March 12, 2026, the student plan changed from “GitHub Copilot Pro” to GitHub Copilot Student*. The 300 Premium Requests/month limit remains, but models like Claude Opus/Sonnet are now only available via Auto mode.
- 2 Antigravity is a cloud-based, agent-first IDE (VSCode fork) released in November 2025. It is not a local tool. Its free-tier rate limits have been reduced multiple times since December 2025 and are currently unstable. Key Point: Use Claude as a “smart advisor” focused on refining specifications, and delegate implementation to free agents.
Overall Coding Workflow
```plain text
[0] Draft a rough specification yourself
↓
[1] Refine the spec with Claude (+ generate prompts)
↓
[2] Delegate implementation to AI agents
↓
[3] Review generated code (test/debug)
↓
[4] If OK, merge PR / commit
↓
Return to [0] and define the next requirement → repeat until complete
---
## [0] Create a Rough Specification
Start by writing down what you want to build in bullet points:
```plain text
- When the user presses a button, the score increases by +1
- The score is displayed at the top of the screen
- When the score reaches 10, transition to a game over screen
Perfection is not required at this stage. As long as you’ve clarified what you want to build, that’s enough.
Writing this on paper can help organize your thoughts. It’s even better if you brainstorm ideas interactively with free AI tools.
[1] Refine the Specification with Claude
Take the rough spec from [0] and pass it to Claude Sonnet / Opus to refine it into a more concrete specification. At the same time, have it generate a prompt for the AI coding agent.
Effective Prompt Template
```plain text
Based on the following specification, create a more detailed specification
and a prompt suitable for an AI coding agent.
If anything is unclear, ask me before implementation.
(Paste your rough specification from [0])
**Important:** Explicitly state “ask me if anything is unclear.”
Otherwise, ambiguity increases the failure rate in later steps.
### Using Claude Code (CLI)
If you’re using Claude Code, open your target repository/folder at this stage. Claude can refine the spec while referencing your codebase.
```plain text
cd <repository path>
claude
Token-Saving Tips
- Opus is powerful but consumes more tokens; Sonnet is often sufficient
- Claude usage depends on output token volume, not request count
- Usage resets every 5 hours → avoid burning through everything at once
- Weekly limits (introduced after August 2025) may apply for heavy users
[2] Delegate Implementation to AI Agents
Use the refined prompt from [1] and send it to agents in this priority order:
Priority
① GitHub Copilot Agent → ② Antigravity → ③ Google Jules → ④ Gemini CLI
① GitHub Copilot Agent
- Type: Cloud (asynchronous)
- Cost: 1 task = 1 Premium Request
- Limit: 300/month (free for students via Copilot Student)
- Success Rate: High
- 특징: Async workflow—submit and wait. Use requests carefully.
② Antigravity
- Type: Cloud-based agent-first IDE
- Cost: Free (preview)
- Features:
- Access to multiple models (Claude Opus 4.6 / Sonnet 4.6 / Gemini 3.1 Pro)
- Planning mode adds a confirmation step before execution
- Manager View allows up to 5 parallel agents
- Notes:
- Rate limits are unstable
- Not a local tool
③ Google Jules
- Type: Cloud (asynchronous)
- Cost: Free (15 tasks/day, max 3 concurrent)
- Model: Gemini 2.5 Pro
- Success Rate: Moderate (~40–50%)
- Features:
- Interactive (Plan) mode allows agent questions
- Good alternative to Copilot Agent
- Supports scheduled tasks (e.g., refactoring, security checks)
④ Gemini CLI
- Type: Local
- Cost: Free (usage-based, replenished daily)
- Success Rate: Moderate
- Features:
- Generous usage limits
- Nano model has low reliability → avoid it
- Prefer Gemini 3 Flash or 2.5 Pro
[3] Review Generated Code
Run and verify the generated code.
If Errors Occur
Copy the error message and send it back to the same agent:
```plain text
The following error occurred. Please fix it.
(Paste error message)
---
### If Behavior Is Incorrect
**Close the PR or discard changes, then return to [2].**
Why not request incremental fixes?
Because repeated patching often breaks other working logic. Resetting and improving the prompt is usually faster and more reliable.
If issues persist, go back to [1] and refine the prompt.
---
### When Manual Fixing Is Faster
If you understand the language, small fixes (e.g., constants) are often quicker to do yourself.
Delegating minor fixes to agents:
- Takes time
- Risks breaking other logic
---
## [4] Merge → Return to [0]
Once everything works, merge or commit and return to [0] for the next requirement.
Repeat this loop until completion.
---
## Tool Comparison Summary
---
## PRO TIPS
### Insert Deep Research Between [1] and [2]
If you get stuck during spec drafting, use Deep Research (Gemini / Copilot / ChatGPT) to explore prior work.
**Steps:**
1. Ask Claude to generate a research prompt
2. Run Deep Research
3. Attach the report to your agent prompt in [2]
This improves coverage and success rates.
---
### Copilot Is Free for Students
GitHub Copilot is available **for free (Copilot Student plan)** via GitHub Education.
You get 300 Premium Requests/month—use it.
---
## Summary
- Use Claude for refining specifications—not implementation
- Delegate coding to free agents (Copilot → Antigravity → Jules → Gemini CLI)
- Reset and retry instead of stacking fixes
- Use Deep Research to improve prompt quality
- Antigravity is a cloud IDE, not local—prepare backups
By using AI tools strategically, you can significantly increase development speed while keeping costs low.
---
*This article was automatically cross-posted from [norinori1's portfolio](https://norinori1-web-site.vercel.app/news/vibe-codings-save-tokens-2026)*
Top comments (0)