Thirty schedules. Three coworkers. One place to look.
Give Your AI a Task Coordinator
When you forget which AI owns the work, give one of your existing coworkers responsibility for the directory.
Thirty scheduled tasks across Gemini Spark, Claude Cowork, and ChatGPT Work create a very practical problem: you cannot remember where you put the work.
Something checks your calendar. Something prepares your weekly report. Something watches for an announcement. You remember asking. You do not remember which AI owns the task, which conversation contains it, or where to change it.
My recommendation is to appoint a Task Coordinator: one existing AI coworker that maintains a durable directory of your scheduled work. Each task continues running with its current provider. The coordinator gives you one place to find it, understand it, and track a requested change.
For students following the frontier-company rule, the boundary is firm: use only OpenAI, Anthropic, Google, and files on your own computer. Do not subscribe to a third-party task manager, automation service, database, or hosting platform. Have the intelligence you already use build this capability with the access you already have.
One directory. One responsible coworker.
For this exercise, choose ChatGPT Work as the coordinator and give it a dedicated project called “Task Coordinator.” This is an organizational choice, not a claim that OpenAI can automatically administer the other providers. Claude Cowork or Gemini Spark can fill the role if either has better access to your chosen directory.
Start with a local project if your existing setup can read and update local files. Ask the AI to create a master directory, a readable dashboard generated from it, and a change log. The master can be a structured JSON file; you should not need to edit that format yourself. Your normal interface is a conversation: “Find my meal plan.”
Give only the coordinator responsibility for updating the master. Other coworkers supply registration receipts. Keeping one writer reduces the chance that competing copies quietly become different directories.
If you need access across devices, use a private Google Sheet in your existing Google account as the master instead, provided the coordinator can actually read and update it through permitted access. Choose one master and record its exact location. Downloaded files and dashboards are dated copies of that master.
Prove persistence before adding automation. Create one test record, close the conversation, and ask a fresh session to retrieve it from the saved location. Then change one field and read it back. A remembered answer is insufficient; the saved record must exist.
Begin with three. Then account for every one.
Build the first version around one existing task from each provider. You can test the whole procedure before entering thirty records. Once it works, complete the inventory across every account and workspace you use, including paused work.
OpenAI
Open Scheduled in ChatGPT to inspect tasks and recent runs. Local project schedules need the computer on and the app running; web tasks cannot directly use a folder on your computer. OpenAI scheduling documentation.
Anthropic
Open Scheduled in Cowork to inspect saved tasks, edit instructions or cadence, and review upcoming and past runs. Check the actual task’s execution environment and available files before relying on unattended access. Anthropic scheduling documentation.
In Gemini, switch to Spark and open Schedules. A task thread can own several schedules. Record each schedule separately and retain its parent task link. Spark schedules are distinct from scheduled actions in Gemini chat. Google scheduling documentation.
Ask the coordinator to inspect the scheduling pages through access it already has. If it cannot reach a page, supply screenshots or copied details. A list screenshot may establish that an entry exists; the task’s details establish its full instructions and timing. Open the details when necessary.
Keep a coverage note: which accounts were checked, which remain unchecked, and when. An inaccessible account is “not yet inventoried.” It is never “zero tasks.” A task discussed in a conversation is only a candidate until its saved schedule is found.
Assign a permanent identifier to each directory record: T001, T002, T003. Add that identifier to editable schedule names when practical. Retain the provider’s own identifier or link separately. Never recycle a retired identifier. If one job has several independent schedules, give each schedule its own record and link the related records.
Store enough to find it—and rebuild it.
The compact view should show purpose, provider, account, timing, status, and an Open Task link. Behind it, keep the details that would let you understand the assignment months later.
- Identity & location
- Permanent ID, exact saved name, plain-language purpose, provider, account or workspace, management link, and originating conversation. If linking is unavailable, save precise navigation steps and a searchable name.
- Instructions & dependencies
- The full saved prompt, referenced skills or files, output destination, and required connections. Store references to credentials, never the credentials themselves.
- Timing & state
- Cadence or event condition, named time zone, next run when available, and the last verified enabled, paused, or completed state. Separate requested changes from saved settings.
- Evidence & freshness
- When settings were checked, how they were checked, the latest observed run or output, and its time. Record unknowns explicitly. A new dashboard build does not refresh old evidence.
For a time-based task, write “Sunday at 4 p.m., America/New_York,” rather than “Sunday afternoon.” For an event monitor, record the actual condition. If the provider follows the device’s time zone, document that behavior instead of assuming the schedule is anchored to a city.
Give time-based deliverables an agreed completion window. A report due Monday morning might warrant investigation after that window. An event monitor needs a different check: silence can mean that no matching event occurred.
Three statements that mean different things.
These examples illustrate the labels your directory should use.
Enabled · Settings verified
The saved schedule is switched on. That establishes configuration. It does not establish that the last run started, finished, or delivered anything.
Completed · Result verified
A run record exists and the expected output was checked. Link to that evidence and show its timestamp. An old successful run says nothing about a later run you have not inspected.
Execution unverified · Follow-up needed
The coordinator cannot inspect the latest execution. Keep the last known settings visible and identify the missing evidence. Do not turn an unknown into either a success or a failure.
A dashboard should save a search.
Once the three-task pilot works, ask the AI to generate a searchable dashboard from the master directory. Use three views: All Tasks, Needs Attention, and Paused. Keep retired records in an archive. Needs Attention should explain the reason: incomplete inventory, stale settings, pending change, missing expected output, or an observed failure.
Search should include purpose, exact name, ID, and simple aliases. Add “dinner,” “meals,” and “groceries” to a meal-planning record. Ordinary keyword matching is enough to begin; the coordinator can interpret broader questions from the saved directory.
A local HTML dashboard is sufficient. It needs no separate hosting account. Have the AI regenerate it after confirmed directory updates and stamp it with its generation time. Show each record’s verification time separately. A beautiful dashboard rebuilt today can still contain last month’s facts.
Keep this personal dashboard private: it contains account context, internal task links, and potentially sensitive instructions. The public article you are reading is a guide to building one; it is not a live view of anyone’s schedules.
Find. Change. Read back. Record.
Suppose you say, “Move my meal plan from Sunday afternoon to Saturday morning.” The coordinator locates T014, identifies its provider and account, and resolves the exact intended time. It then changes the original schedule through authorized access, reopens the saved setting, and updates the directory from that evidence.
If it cannot make the change, it supplies the exact link and handoff instruction. The record keeps the last verified schedule and adds “Change pending: Saturday at 9 a.m.” The requested time must not replace the verified time until the original confirms it.
Apply the same procedure to pausing and retiring tasks. Preserve the instructions and history of retired work. A task that disappears from a page should first be investigated: it may have moved accounts, become inaccessible, or been deleted.
When moving work between providers, keep the permanent directory ID and add the replacement’s provider link. Prepare the replacement without an active recurring schedule, verify its instructions and dependencies, pause the original and read that state back, then enable the replacement. Inspect its first scheduled result. Record any temporary coverage gap; avoid two active owners for the same job unless duplication is intentional.
Registration first. Reconciliation second.
Add this rule to the persistent instructions available in each coworker: whenever a scheduled task is created, changed, paused, or retired, update the coordinator or produce a registration receipt.
Task ID, if assigned. Exact name. Provider and account. Management link. Schedule or trigger and time zone. Saved state. Output destination. What changed. When and how the saved result was verified.
New receipts without an ID go to the coordinator for numbering. The coordinator checks whether the provider link already exists before creating another record. This prevents one schedule from becoming two directory entries.
Then establish a weekly reconciliation: compare the master with the actual scheduling pages and investigate missing entries, changed times, and overlapping purposes. Registration captures known changes quickly. Reconciliation catches changes that escaped registration.
Test reconciliation manually before scheduling it. The background coordinator must reach both the master directory and the relevant scheduling evidence in its actual execution environment. Foreground browser access does not prove background access. Where access is missing, schedule a request for fresh details and label those records as awaiting verification.
Register the coordinator’s own maintenance task. Display its last attempted refresh and last successful reconciliation. A failed refresh must leave the previous verified records intact. Ask for notifications only when something changes meaningfully or needs your action, using an existing OpenAI, Anthropic, or Google channel.
Build my Task Coordinator.
I have scheduled work across ChatGPT Work, Claude Cowork, and Gemini Spark. Build a Task Coordinator in this project using only my existing OpenAI, Anthropic, and Google tools and storage, or local files. Do not recommend, register for, subscribe to, or depend on third-party services.
Choose one durable master directory and state its exact location. You are its responsible writer. Start with one real schedule from each provider. Prove that a fresh session can read the saved records and that an update persists.
Record each schedule separately with a permanent ID, purpose, exact name, provider, account, management link or navigation instructions, full saved instructions, dependencies, schedule or trigger, time zone, output destination, verified state, and dated evidence. Identify unchecked accounts and missing access. Do not invent records from conversational memory.
Generate a searchable dashboard with All Tasks, Needs Attention, and Paused views. Keep archived records. Show dashboard generation time separately from settings verification and execution evidence. Label examples, unknowns, stale records, and pending changes clearly.
Change the original schedule only through available authorized access, read back the saved result, and then update the directory. Otherwise provide a precise handoff and preserve the last verified state. Maintain a change log and a recoverable previous copy of the master.
After the pilot passes, complete the inventory. Establish registration receipts and a weekly reconciliation routine, including a record for the coordinator itself. Test background access before promising automatic reconciliation. Notify me through an existing permitted channel when action is needed.
The test is simple.
Find three real tasks through the directory. Open each original. Make one schedule change, read it back, and confirm that the directory agrees. Reopen the saved directory from a fresh session.
Then you have a dependable answer to “Where did I put that?” The intelligence you already use has taken responsibility for keeping the map.