The Task template makes your AI handle any work item with the same six steps: confirm a title and summary, write a one-screen plan you approve, implement it surgically, verify against the plan, and capture anything non-obvious it learned. You approve twice, and a failed verification returns to Implement.
6 steps · 2 approvals · 1 loop back
example · the steps are this template's real steps
For anyone who hands everyday work to an AI (refactors, doc updates, research) and wants a plan before the change and a check after it.
It is a stored workflow, not a prompt. Your AI client receives one step at a time from ConvOps over MCP, does the work, and advances. The engine holds the run at each of its two approvals until a person says yes.
6 steps, in this order. Below each one is the instruction your AI receives at that step, exactly as the template stores it.
Confirm the task has a clear title and a one-line summary. If either is missing, pause and ask the operator. Do not advance until both are present.
Write a short plan: the minimum needed to know what to build and what could go wrong. One screen of text, not a document.
Cover:
Ask the operator to approve the plan before moving on.
no rejection path · the run moves on only when approved
Execute the plan from the previous step. Keep changes surgical: don't scope-creep into adjacent code.
When finished, move to verify.
Run the verification you defined in the plan. Report what you ran, what passed, what didn't. If something failed, go back to implement and fix it before asking for approval.
Ask the operator to approve before moving on.
rejected → back to Implement
If anything non-obvious came up, write a short note to:
docs/learnings/{YYYY-MM-DD}-{task-slug}.md
Use plain Markdown. Skip if nothing non-obvious happened.
Task done. Close it out.
Two steps wait for a person. You answer in the chat; the engine records the verdict and the audit row records who gave it.
The plan fits on one screen: what and why, files likely to change, obvious risks and how it will be verified. You approve it before anything changes.
what you review
no rejection path · approve to continue
Your AI runs the verification it promised in the plan and reports the result. Reject it and the run returns to Implement.
reject → back to Implement
gates the engine checks without you
Read from the template's own steps: the tools they call, the files they write, and whether they touch code.
Connect your AI client first, then install Task from the catalog in the ConvOps app. See how to connect a client
An installed template is a workflow in your workspace. Ask your AI to change it in plain words; the edit lands on the one step that needs it.
new chat
more example prompts
Task plans first and implements against that plan. Bug Fix reproduces and finds a root cause before any fix, and requires a regression test. Task attaches to the task type, Bug Fix to the bug type.
Capture Learnings writes a short Markdown note to docs/learnings/{YYYY-MM-DD}-{task-slug}.md when something non-obvious came up, and skips the step otherwise. The closing note your AI leaves at Complete is saved to the task as context.
No. The steps are generic: intake, a plan, the work, a check. The catalog examples include refactoring an auth module, updating API docs and researching competitor pricing.
Any MCP client connected to ConvOps. Verified clients: Claude Code, Claude (desktop, web, Cowork), OpenAI Codex, ChatGPT, Cursor, OpenCode. Because its steps change code, run it in a coding client with your repository checked out.