EndlessZenFramework

A minimal starter for one autonomous researcher in ChatGPT, with GitHub as a research source and optional publication surface, and Google Drive as persistent project memory. It contains instructions and templates, not a running agent, an installer, or a guarantee of uninterrupted execution.

Set up one researcher

1. Define a narrow research question, intended reader, deliverable and completion gates. Name the repositories the worker may access. GitHub is read-only by default; publication needs an explicit grant for each repository and action.
2. Connect GitHub and Google Drive in the ChatGPT experience you will actually use. Check its available actions and permissions, not just whether an app is connected. See the official app guidance and GitHub guidance. A read-only connection is enough for repository research, but not for publishing issues or PRs.
3. Create one private Drive project folder. Copy and complete PROJECT_MANIFEST and STATE_AND_QUEUE there. Keep their stable file IDs in the completed prompt. Create an Artifacts subfolder only for the first result.
4. Copy the researcher prompt into a dedicated chat, replace its placeholders, and run one supervised cycle. Authorize a small Drive create/update/readback trial inside the project folder. Verify the file's parent and contents. Do not assume the same capabilities exist in scheduled runs.
5. Using the completed prompt, request one recurring task at an available cadence in your timezone. Keep the queue in Drive, not in the task prompt. The first scheduled cycle verifies its own access and write ownership before research. Mark setup ready only after that evidence is recorded and read back.

Example scheduling request, after completing the private setup:

Run the following researcher instructions once per hour in <IANA timezone>.
Report briefly after every run, including blocked and no-work outcomes.
<PASTE THE COMPLETED RESEARCHER PROMPT HERE>

The cadence must be supported by the account. Scheduling and app availability depend on the selected experience and permissions; some actions may require approval. Read the current task documentation. Do not rely on chat history or uploaded project files as the scheduled worker's control state.

Minimum operating contract

Every wake reads durable state, recovers interrupted work, and performs one bounded unit: ORIENT -> SELECT -> CLAIM -> PRODUCE -> VERIFY -> COMMIT. Research ends in a supported decision or an explicit inconclusive result. Producing an answer and accepting it happen in different runs. GitHub facts are read from the live repository; Drive owns the research queue, evidence and decisions.

Safe writes require verified serialization of all entry points, including manual runs and retries, or enforced conditional ownership covering every affected write. One timer, a claim field or a revision counter is not a lock. This repository supplies no lock implementation. Until the selected environment proves safe ownership, keep writes supervised under verified exclusion, or remain read-only and report the gate.

A worker must not stop its own recurring task because it is blocked, finished or out of work. The owner manages the schedule and checks for missing run reports; a worker cannot report a wake that never happened. Provider-side pauses remain possible.

Keep deployment data private

Only blank templates belong here. Keep filled prompts, Drive URLs and IDs, control state, research notes, account details, credentials and run history outside this public repository. Publish only authorized, sanitized findings that a reader can understand without private sources. Never import another repository's Git history.

The entire starter is this README, the prompt, two control templates, contributor instructions in AGENTS.md, and the original MIT license. There are no bots, product integrations, execution controllers or code dependencies.