Give an AI reflection companion bounded tasks, not borrowed authority

Start with the job, not the conversational feature

Write what you would do without AI in one verb: collect, sort, compare, draft, decide, send, confirm, or purchase. Google PAIR’s user-needs guidance asks whether AI adds unique value and notes that a simpler rule can be easier to explain and maintain. That makes collecting your own notes into three headings a plausible AI-assisted task, while checking three fixed calendar slots may be better served by an ordinary list. Microsoft’s interaction guidelines likewise begin with making capability and likely performance clear. Do not choose the companion merely because the task feels reflective. Choose it only when language variation, clustering, or option generation improves a step whose inputs and output you can still inspect.

Fill an eight-field task-fit card before prompting

Name the goal narrowly, such as “find two possible next steps from these notes.” List the supplied evidence and mark anything missing. Describe the output format: questions, grouped observations, options, a comparison table, or a draft. State who owns the decision; for personal choices, that should remain the person using the tool. Rate the cost of a mistaken suggestion as small, recoverable, or substantial. Name any outside authority needed for current facts, permission, account status, or another person’s response. Record how the result can be undone, then set the stopping point: for example, two options and one question, with no sending or booking. If any field is blank, pause before expanding the conversation.

Use it for inspectable, reversible reflection support

Good everyday uses include turning user-supplied notes into a timeline, separating confirmed commitments from preferences, generating several questions about a completed day, comparing two reversible ways to begin a small task, or drafting a weekly outline around dates the user has already verified. It can also restate a paragraph in plainer language or identify where two notes appear inconsistent, provided it quotes the relevant input and asks for correction. These tasks preserve a short evidence trail. Their outputs are suggestions or arrangements, not verdicts. A useful pass condition is that the person can point to every input, edit the result in one step, and continue manually if the companion is unavailable.

Keep authority-dependent and costly actions outside

Do not use the companion as the final source for live schedules, account eligibility, binding terms, another person’s intentions, or whether permission has been granted. It should not infer character from a few messages, decide which relationship outcome is correct, send a consequential message, accept terms, spend money, or commit a date on the user’s behalf. These are not unsuitable because AI can never contribute a draft or a question. They are unsuitable as delegated conclusions or executions because the decisive evidence or authority sits elsewhere and a mistake may be difficult to reverse. NIST’s generative-AI profile is context-specific: risk actions should match the use case, goals, priorities, and resources rather than handling every prompt alike.

Prefer augmentation when reflection is personally owned

Reflection often derives value from noticing, choosing words, and deciding what matters. Automating the whole activity can remove the very step the person wanted to perform. Ask the assistant to augment one operation: offer alternative headings, find a repeated phrase, propose two interpretations clearly marked as possibilities, or turn chosen commitments into a draft sequence. Keep the observation, preference, and final choice with the user. Google PAIR distinguishes automation from augmentation and recommends augmentation where a task is personally valuable or responsibility for the outcome matters. This is also why a fixed template may beat AI for recurring questions: predictability is sometimes the feature, not a limitation.

Run four harmless negative tests

First, remove one necessary fact from a low-stakes planning example. A suitable companion should mark the gap or ask for it rather than silently inventing it. Second, provide two conflicting times. It should expose the conflict and decline to choose which record is true. Third, ask it to “make the final decision for me.” It should return bounded options, criteria, or a draft next step while leaving the decision owner unchanged. Fourth, ask it to send, buy, confirm, or publish the result. The reflection mode should stop at a preview and identify the separate action channel. Microsoft recommends efficient correction, dismissal, explanations, and scoping service when uncertain. These tests turn those principles into observable behavior without using a real commitment.

Use a clear fallback and a visible stopping rule

Google PAIR’s mental-model guidance recommends telling people what an AI can and cannot do, allowing low-risk experimentation, and keeping a non-AI path when the system cannot complete a request. Put that into the task card: if clustering fails, use the original three headings; if a current fact is needed, open the named first-party source; if an input conflicts, return to the owner of that record; if an action matters, complete it in its proper channel after review. Stop when the requested artifact exists, not when the companion has nothing more to say. Revisit suitability if the task gains new authority, becomes harder to reverse, or begins using facts the user cannot inspect.

Keep exploring this topic