# AI Task Brief: From Problem to Human Acceptance Copy this into a private project directory before a model call. It is not a prompt collection. It is permission to start: when task, data, sources, acceptance, or ownership remain unclear, do not ask a model for the final result yet. Do not include passwords, identity documents, precise addresses, medical privacy, children's data, customer records, unpublished vulnerabilities, or unauthorised third-party material. Redact sensitive material or use an organisation-approved environment. ## 1. Task and Ownership ### AI Task Brief - YYYY-MM-DD Real situation:\ User/audience:\ Decision or action to complete:\ Why now:\ Deadline and non-negotiable checkpoints:\ Final owner and human reviewer:\ Who is affected if this is delayed or not done: Replace “learn AI” or “build an intelligent assistant” with an observable action, such as “a user can import a file in ten minutes and see an explainable report with sources and errors”. ## 2. Inputs, Sources, and Data Boundary | Input/claim | Type | Source, version, location | May reach model? | Verifier | Expiry condition | | ----------- | ------------------------------------------------------- | ------------------------- | ----------------------------------------- | -------- | ---------------- | | | fact / inference / experience / user data / third-party | | public / redacted / approved / prohibited | | | Files and fields allowed:\ Material deliberately withheld:\ Data sensitivity: public / internal / confidential / restricted\ Collection, transfer, retention, and deletion dates:\ Required consent obtained:\ Copyright, licence, citation, and authorship requirements:\ Minimum data scope the model can see: A source is not automatically true. Each critical claim must return to an original passage, version, data definition, or direct observation. Model links, numbers, and quotations still require human verification. ## 3. Output and Evaluation Set Final deliverable:\ Format, length, and audience:\ Observable definition of done:\ Hard gates that must pass:\ Acceptable variation:\ Content that must not appear: Build a small, realistic evaluation set instead of only model-friendly examples: | Sample | Input condition | Expected result | Unacceptable result | Actual result | Evidence location | | ------------------------- | --------------- | --------------- | ------------------- | ------------- | ----------------- | | Normal case | | | | | | | Boundary case | | | | | | | Missing/conflicting input | | | | | | | Redacted historical case | | | | | | Acceptance should answer whether facts are traceable, the action is complete, errors are visible, permissions are correct, and failure can stop. Fluent, fast, or human-like is not an acceptance standard. ## 4. AI's Working Scope AI may: ask / classify / propose explanations / give counterexamples / transcribe\ / draft / suggest tests\ AI may not: make final factual decisions / invent sources / approve for the owner\ / cross permissions / complete a prohibited exam or application\ / publish or execute automatically\ Mode: diagnosis / assistance / candidate generation / batch processing / other\ Model, version, region, and call date:\ External tools, MCP servers, or other protocols and versions:\ Read/write scope, authorizer, and revocation method for each tool:\ Prompt, system instruction, or workflow version location: For a long-running or tool-using task, also record: Run/task ID and parent task ID:\ Async or synchronous mode; expected and maximum wall-clock time:\ Checkpoint locations and resume procedure:\ Maximum retries, timeout, and idempotency key:\ Browser pages, screenshots, and external instructions treated as untrusted input:\ Network egress allowlist and credential scope:\ External writes, receipts, and final environment state:\ Human takeover point and emergency stop:\ Supplier policy/version, effective date, region, and high-risk review: Ask the model to restate goal, input, limits, unknowns, and acceptance before generation. Version prompts, and never treat one chat window as the project's only record. ## 5. Human Gates | Gate | Who confirms | Passing evidence | If it fails | | ----------------------- | ------------ | ------------------------------------------------------------------- | ------------------------------------ | | Source gate | | Original link, version, location | Mark unverified; do not circulate | | Fact gate | | Sample checks and data definition | Delete or downgrade claim | | Privacy/permission gate | | Scope, consent, access record | Stop and redact | | Quality gate | | Evaluation set, edge cases, real feedback | Repair, narrow, or reject | | Cost gate | | Tokens, time, human rework, budget | Downgrade or stop | | Tool authorization gate | | Tool scope, call record, named authorization, and revocation method | Block the call or return to a person | | Ownership gate | | Named approval and disclosure | Do not publish or execute | Critical decisions cannot be approved only by the generator, an automated score, or one developer alone. High-risk domains return to current primary sources and qualified professionals. ## 6. Cost, Retention, and Reversibility Call volume, time, and cost ceiling:\ Human review and rework budget:\ Does data enter training, logs, or third-party retention:\ Acceptable latency and downgrade path:\ Version that can be withdrawn, rerun, or restored:\ Release scope and pilot audience: A cheap call that causes a privacy incident, wrong decision, or major rework is not cheap in reality. Record human time, review, failure, reruns, and communication in addition to model fees. ## 7. Failure, Pause, and Rollback | Trigger | Immediate action | Notify | Recovery/rollback location | | --------------------------------- | ------------------------------------ | ------ | -------------------------- | | Source or version cannot be found | Stop circulation; return to original | | | | Evaluation hard gate fails | Block release or automatic action | | | | Input crosses permission boundary | Stop upload; revoke access | | | | Cost/latency exceeds ceiling | Downgrade, throttle, or stop | | | | Real user reports harm | Remove, preserve evidence, escalate | | | Without a named stop owner, notification path, and known-good version, the task is not ready for a real workflow. ## 8. Handover and Public Disclosure Current state: not started / experiment / internal pilot / limited release / delivered / stopped\ Completed work and evidence locations:\ Open questions and risks:\ Smallest next task:\ Handover owner, date, and access:\ What readers/users need to know about AI involvement:\ Source, privacy, copyright, and conflict-of-interest note: The next operator should not search chat history to discover what to do. Public work should state which step used a tool, which facts a person confirmed, and which content remains a candidate. ## 9. Preflight Check - [ ] Audience, action, deadline, and owner are explicit. - [ ] Input fields, sensitivity, consent, copyright, and retention are confirmed. - [ ] Critical sources, versions, and expiry conditions are traceable. - [ ] Normal, boundary, conflict, and redacted historical cases are in the evaluation set. - [ ] AI may/may-not scope is explicit. - [ ] Long-running checkpoints, retry/idempotency rules, browser-input boundary, and network egress are recorded when applicable. - [ ] Human gates, cost ceiling, stop conditions, and rollback location are named. - [ ] Handover, disclosure, and next review date are set. --- Source: [Life Level-up Guide](https://byoungd.github.io/up/en/templates/ai-task-brief) · Manuscript updated: 2026-10-16. Content: [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).