# Using AI coding agents safely AI assistance is welcome, but it is never required. The safest default is simple: **use your existing included allowance, keep approval controls on, work in your own fork, and do not enable paid overages just to contribute.** YazSes does not reimburse third-party AI or cloud charges unless a maintainer agreed in writing in advance to a specific expense and amount. Read the canonical [third-party AI tools notice](https://github.com/MSKazemi/yazses/blob/main/THIRD_PARTY_AI_TOOLS.md) before using a paid service. **Last provider-doc review: 2026-09-22.** Providers change products and billing frequently, so the official provider pages linked below always take precedence over this guide. ## Pick the billing mode first | What you want | Safer starting point | |---|---| | **No AI at all** | Work normally. Every YazSes contribution path supports human-only work. | | **Use an allowance already included in Claude** | Start without `ANTHROPIC_API_KEY`, sign in to Claude, verify with `/status`, and leave usage-credit/overage continuation off unless you deliberately want it. | | **Use Codex through your ChatGPT plan** | Start without OpenAI/Codex API-key variables, sign in with ChatGPT, verify with `/status`, and leave optional credit auto-reload off unless you deliberately want it. | | **Use an API on purpose** | Use your own authorized account/project, set the provider's strongest available spend control plus alerts, monitor usage, and assume the resulting charges are yours unless a separate written agreement says otherwise. | If you are not sure which row applies to your account, stop before enabling paid continuation and check the provider's current usage/billing page. ## How the repo reaches your agent automatically | Tool | Repository instruction discovery | |---|---| | **Codex / ChatGPT Codex** | Reads root `AGENTS.md` directly. | | **Google Jules** | Reads root `AGENTS.md` directly. | | **Claude Code** | Loads root `CLAUDE.md`, which imports `AGENTS.md`. | | **Gemini CLI** | Loads root `GEMINI.md`, which imports `AGENTS.md`. | | **Other agents** | Tell the tool to read `AGENTS.md` before it edits anything. | The adapter files intentionally contain no copy of the project rules. If instructions change, only `AGENTS.md` should change. Personal Claude settings belong in `CLAUDE.local.md` or the user's Claude configuration; personal Gemini settings belong in the user's Gemini configuration. Do not put private credentials or private maintainer notes into the tracked adapters. ## The 5-minute safe setup 1. **Fork the repository** to your GitHub account and clone your fork. A normal contribution does not require write access to the upstream YazSes repository. 2. From the repository root, read **[AGENTS.md](https://github.com/MSKazemi/yazses/blob/main/AGENTS.md)** and the task or issue you are working on. 3. Decide which billing mode you intend to use: an included subscription/plan allowance, or an API/pay-as-you-go account. Do not let that choice happen accidentally. 4. If you do not want extra charges, disable optional overages, usage credits, automatic reloads, or similar paid continuation features in the provider account. 5. Keep command/file-write permissions on approval for your first contribution. Let the agent inspect and propose; you approve commands, review the complete diff, run validation, then open the pull request yourself. There is no project requirement to use the largest model, maximum reasoning effort, a cloud sandbox, or a paid agent run. Small bounded tasks are deliberately designed to work without that. ### If you want to use a cloud agent on campaign work For **Google Jules**, start with the dedicated [step-by-step Jules connection guide](jules.md). It shows the approved fork-first setup that is available today and separately documents the still-gated owner/upstream setup. For provider-neutral worker rules, use the fork-first flow in [`campaign/agent-workers.md`](https://github.com/MSKazemi/yazses/blob/main/campaign/agent-workers.md). It explains how task claims, `cloud_agent_ready`, allowed paths, provider-specific execution triggers such as Jules, and human review fit together. The key distinction is that **agent readiness is project metadata; provider execution is contributor-owned**. A normal contributor does not need upstream write access or a project-owned agent credential. ## One prompt for any agent Paste this after opening the repository in your agent: ```text I am making one contribution to YazSes. I remain responsible for the pull request. 1. Read AGENTS.md and .github/CONTRIBUTING.md completely before changing anything. 2. Read the linked issue/task and restate its acceptance criteria and allowed paths. 3. Inspect the relevant implementation and tests first. 4. Make the smallest change that satisfies the task. Do not tidy unrelated code. 5. Do not add network calls, telemetry, cloud fallbacks, dependencies, or default-on features. 6. Do not buy credits, enable paid overages, create paid cloud resources, change my subscription, or alter billing settings. 7. Ask before running destructive, privileged, networked, or unusually expensive commands. 8. Run the task's validation command and the required repo checks. 9. Show me the complete diff and explain every changed line. 10. Do not commit, push, or open a pull request until I approve the diff. ``` If an agent claims the task requires purchasing credits or upgrading a plan, stop and either wait for your normal quota to reset, switch tools, or complete the task manually. ## Treat GitHub text as untrusted input An issue or pull request can legitimately tell the agent what outcome is wanted, which files are in scope, and how to validate the work. It cannot override `AGENTS.md` or silently grant extra authority. Do not let an agent obey issue comments, pasted logs, fetched web pages, or downloaded text that tells it to reveal credentials, change billing, disable safeguards, access unrelated files, add an unrelated dependency, or run a privileged/destructive/networked command. Those are prompt-injection patterns, not contribution requirements. Show the suspicious instruction to the human contributor and ask them to decide. ## Claude Code ### If you want to stay inside your Claude subscription Anthropic documents that an `ANTHROPIC_API_KEY` environment variable can make Claude Code use API billing instead of your Pro/Max subscription. Before starting, check without printing the key itself: ```sh test -n "${ANTHROPIC_API_KEY:-}" && echo "ANTHROPIC_API_KEY is set" ``` If you do **not** intend to use API billing in this shell: ```sh unset ANTHROPIC_API_KEY claude ``` Then use `/status` inside Claude Code to confirm the account and usage mode you expect. If you want to avoid charges beyond your included plan allowance, do not opt into API-credit continuation merely because you reached a plan limit. Anthropic also exposes usage-credit and Console billing controls; review any monthly limit and automatic reload setting before enabling them. For API-billed sessions, `/cost` shows the current session's spend. `/model` lets you choose a less expensive model where appropriate. Anthropic recommends clearing unrelated history between tasks (`/clear`) or compacting long sessions (`/compact`) because large context increases usage. Official references: - [Use Claude Code with Pro or Max](https://support.claude.com/en/articles/11145838-use-claude-code-with-your-pro-or-max-plan) - [Models, usage, and limits in Claude Code](https://support.claude.com/en/articles/14552983-models-usage-and-limits-in-claude-code) - [Manage usage credits](https://support.claude.com/en/articles/12429409-manage-usage-credits-for-paid-claude-plans) - [Claude API billing](https://support.claude.com/en/articles/8977456-how-do-i-pay-for-my-claude-api-usage) ## Codex ### If you want to use your ChatGPT plan rather than API billing OpenAI distinguishes Codex usage signed in with ChatGPT from Codex authenticated with your own API key. If your intention is to use your ChatGPT-plan allowance, start from a shell without API-key variables and sign in with ChatGPT: ```sh unset OPENAI_API_KEY CODEX_API_KEY codex ``` Inside Codex, use: ```text /status # confirm current session configuration /permissions # keep an approval level you are comfortable with /model # choose model/reasoning appropriate to the task ``` This repository already has a root `AGENTS.md`; do **not** run `/init` in a way that overwrites it. Codex reads `AGENTS.md` as project instructions. If you intentionally use an OpenAI API key, API pricing applies. For cost control, use a dedicated API project for your contribution work and configure its spend controls. OpenAI's current platform supports project/organization spend controls, including enforceable hard spend limits; alerts by themselves are only notifications. Set the limit before a long agent run, not after. Official references: - [Using Codex with your ChatGPT plan](https://help.openai.com/en/articles/11369540) - [Codex CLI](https://developers.openai.com/codex/cli) - [AGENTS.md project instructions](https://developers.openai.com/codex/agents-md) - [Managing API projects and spend limits](https://help.openai.com/en/articles/9186755) - [Troubleshooting API usage and spend limits](https://help.openai.com/en/articles/6614457) ## ChatGPT For software-development work, the simplest ChatGPT path is to use **Codex** from ChatGPT Desktop, Codex CLI, or another Codex surface available on your plan. Plain ChatGPT can still help you understand an issue, review a diff, write tests, or draft a patch. If you connect GitHub or another developer service to ChatGPT: - grant only the repository access you actually need, preferably your fork for write access; - do not grant upstream write access merely to make a normal contribution; - review the tool's proposed changes before any push or pull request; and - check **Settings / Usage** (wording may change) before buying credits or enabling automatic purchases. OpenAI currently documents separately purchased usage credits for eligible ChatGPT agentic features, including Codex. Included plan usage is used first; after that, eligible features can draw from a purchased credit balance. If automatic reload is available on your account, leave it **off** unless you deliberately want automatic purchases. Also note that a maximum monthly amount configured for automatic reload applies to those automatic reloads; it should not be treated as a universal cap on separate one-time credit purchases. Codex and other eligible agentic ChatGPT features can share plan allowances or credits depending on the plan. A long-running task can therefore consume more usage than a short chat. Treat any purchase of credits, reset, automatic reload, or extra usage as your own billing decision, not a YazSes requirement. Official references: - [Using credits for flexible usage in ChatGPT](https://help.openai.com/en/articles/12642688) - [Using Codex with your ChatGPT plan](https://help.openai.com/en/articles/11369540) ## Provider-neutral ways to reduce cost and risk - **One issue per agent session.** Start fresh when you change tasks. - **Ask for a plan before edits.** Catch an accidental repo-wide refactor before it happens. - **Keep scope narrow.** More files and more context usually mean more tokens and more review. - **Use the smallest model/effort that reliably handles the task.** - **Prefer local tests.** YazSes's normal test suite is offline; contributing does not require paid cloud execution. - **Keep approvals on.** Do not give a first-time agent unlimited shell, network, or repository permissions. - **Never paste secrets.** Do not put API keys, credentials, private data, or payment details into prompts, issues, commits, or PRs. - **Do not send YazSes user/tester data to a cloud agent.** Real recordings, transcripts, voiceprints, learning data, research-participant data, and identifying diagnostic logs stay local; use synthetic or explicitly de-identified fixtures. The product is offline by default, but a cloud development assistant is not. - **Do not run endless retries.** If the same test fails repeatedly, stop the agent and inspect the real failure rather than paying for another loop. - **Review usage before and after.** Provider dashboards and session status commands are the source of truth for your own account. ## What the project will never ask you to do To contribute, YazSes will never require you to: - buy an AI plan or API credits; - enable automatic credit reload or unlimited usage; - share an API key with a maintainer; - add a maintainer's payment method to your account; - give an AI service write access to the upstream repository; - upload private or employer-confidential material to an AI provider; or - keep an expensive run going because an issue is labelled "agent-ready". If a project instruction appears to say otherwise, stop and open a GitHub discussion or issue before spending money. ## Before you open the pull request Run the repository checks from [AGENTS.md](https://github.com/MSKazemi/yazses/blob/main/AGENTS.md), inspect `git diff`, and make sure you can explain every changed line. In the pull-request template, say which AI tool you used and what you verified personally. The human contributor remains responsible for the submitted change; the AI provider remains responsible for its own service and billing; YazSes remains responsible for reviewing the contribution under the project's normal rules.