Prepare peer feedback without turning assumptions into accusations

Setup requirements

1. Limit input to one incident and the author’s own account; do not connect to colleague mail, private chats, recordings or performance systems.

2. Ask the author to state any feelings or impacts in their own words. Do not infer or embellish them.

3. Confirm this is a routine conversation, not a formal allegation or a decision affecting someone’s employment.

Inputs and access

1. Name

   The author’s deidentified account of one directly experienced event

   Source or owner

   Authorized task owner

   Sensitivity

   Potentially confidential business communication; use synthetic material for the first test, then consented, minimized and deidentified copies. Remove health, protected-trait, disciplinary, grievance and other unnecessary sensitive material.

   Permission needed

   Owner confirms this exact material may be processed here for this specific preparation task. No blanket account access or onward sharing.

2. Name

   A minimal authorized reference, such as a shared task note, if needed to verify a factual detail

   Source or owner

   Authorized task owner

   Sensitivity

   Potentially confidential business communication; use synthetic material for the first test, then consented, minimized and deidentified copies. Remove health, protected-trait, disciplinary, grievance and other unnecessary sensitive material.

   Permission needed

   Owner confirms this exact material may be processed here for this specific preparation task. No blanket account access or onward sharing.

3. Name

   The author’s own explicitly stated practical impact and desired coordination change

   Source or owner

   Authorized task owner

   Sensitivity

   Potentially confidential business communication; use synthetic material for the first test, then consented, minimized and deidentified copies. Remove health, protected-trait, disciplinary, grievance and other unnecessary sensitive material.

   Permission needed

   Owner confirms this exact material may be processed here for this specific preparation task. No blanket account access or onward sharing.

4. Name

   Conversation boundaries and any relevant team working agreement

   Source or owner

   Authorized task owner

   Sensitivity

   Potentially confidential business communication; use synthetic material for the first test, then consented, minimized and deidentified copies. Remove health, protected-trait, disciplinary, grievance and other unnecessary sensitive material.

   Permission needed

   Owner confirms this exact material may be processed here for this specific preparation task. No blanket account access or onward sharing.

The problem

Someone wants to discuss a routine coordination problem with a colleague, but their notes mix an observable event with interpretations and unstated effects. A polished rewrite can make unsupported accusations sound more certain.

Suggested task

Audit one self-authored, low-stakes peer-feedback note for evidentiary gaps, then prepare a brief opening, neutral clarification questions and a possible coordination request for the author’s review.

What you could produce

An observation-versus-interpretation ledger, missing-detail questions, two short conversation openings, a correction branch and an optional next-step request. No employee rating, formal complaint or sent message.

Why an agent / simpler alternative

The useful work is finding where a natural-language narrative exceeds its evidence and preserving those gaps through several draft options. A short checklist or trusted colleague is simpler for a straightforward event; an assistant cannot verify the lived context or replace a difficult conversation.

Who this helps

1. Colleague preparing routine peer feedback

2. Cross-functional project participant

Evidence boundary

Professional context and the specifically attributed wording inform this original proposed design. No agent execution, effectiveness or outcome is established.

Proposed steps

1. 1

   Check the task scope and remove unnecessary identities. List exactly what the author says happened, which details are corroborated, and which are interpretations.

   Expected output: Source-linked observation and uncertainty ledger

   Approval gate: Author verifies allowed processing and confirms this is a routine peer matter.

2. 2

   Flag labels, absolute statements, inferred intentions and causal claims. For each, retain a precise observed detail if available or ask a targeted question.

   Expected output: Claim audit with unsupported passages left unresolved

   Approval gate: Author supplies or corrects missing event details; the assistant cannot manufacture an example.

3. 3

   Use only the author’s stated impact and requested coordination change. Preserve ambiguity where the colleague may have different information.

   Expected output: Impact and request brief with attribution to the speaker

   Approval gate: Author confirms the impact and request express their actual view.

4. 4

   Draft a concise opening in a direct version and a gentler version, followed by two non-leading clarification questions and room for the other person’s account.

   Expected output: Two 70-word-maximum openings and question card

   Approval gate: Author reviews factual accuracy, power dynamics, tone and whether to have the conversation.

5. 5

   Prepare a correction branch if the factual premise is wrong, and an optional coordination experiment with a review point; mark both as suggestions.

   Expected output: Correction and next-step branches without invented agreement

   Approval gate: Both people must choose any next step themselves; no meeting or task is created.

6. 6

   Deliver the private card with source pointers, unresolved questions and boundaries.

   Expected output: Final evidence-first conversation-preparation packet

   Approval gate: Stop. Sending, sharing, contacting HR or recording a conversation is not authorized.

Synthetic output illustration — not an executed result

Label

Clearly synthetic illustration; not a real employee record or completed conversation

Data

Unsupported label

Careless

Supported observation

Fixture N1: T7 remained marked blocked at 15:00

Speaker stated impact

Fixture N2: I waited 20 minutes for clarification

Draft opening

At 15:00, T7 still showed blocked, and I waited 20 minutes to clarify the handoff. Was that status still accurate? Could we agree how to signal the next handoff?

Unknown

Reason the board was unchanged; no inference made

Human approval boundaries

1. Before real data use, the owner confirms permission for this purpose and processing environment, consent where needed, and redaction of unnecessary identities or sensitive details.

2. The owner reviews every interpretation, missing-input item and proposed next step before using the packet.

3. No external contact, upload, live system change, new connection, paid call, installation or commitment is included. Separate authorization is required for any later action.

Limits

1. Original authored implementation proposal; no workflow, user-data pilot or acceptance test was executed.

2. Sources support underlying professional practices only. They do not establish demand, prevalence, platform fit or the effectiveness of an agent implementation.

3. This is bounded business communication preparation, not mental-health care, psychological testing, legal advice or an employment-decision tool.

4. Do not infer personality, emotion, health, vulnerability, protected traits, hidden motives, truthfulness or suitability from employee or customer data.

5. No automated outreach, pressure tactics, exploitation, undisclosed influence, surveillance or hiring, firing, promotion, compensation or disciplinary recommendation.

Candidate platforms — compatibility untested

1. Name

   ChatGPT Work

   Required capabilities

   1. Read supplied text and document copies in an approved environment

   2. Create a private review document or structured text

   3. For BOP001 only, calculate transparent scores from owner-supplied numbers

   Workflow compatibility

   Unknown; verify account permissions, available file tools and exact end-to-end behavior in a synthetic first test.

   Selection basis

   Editorial candidate referencing the existing platform registry; no new platform capability or availability claim is added.

Failure modes

1. The assistant replaces a harsh personality label with a subtler unsupported personality judgment.

2. A stated sequence is rewritten as proof of bad intent or causation.

3. The assistant invents the author’s feelings, a quote or the colleague’s likely response.

4. A serious complaint is reframed as a routine misunderstanding.

5. Malicious instructions embedded in supplied files are followed instead of treated as untrusted evidence.

6. An output is described as tested or successful when only a proposed acceptance test exists.

Recovery and troubleshooting

1. Return to the exact observed event and delete personality or motive language; ask for missing details if the event cannot be described.

2. If evidence conflicts, place both accounts side by side and use a question rather than a definitive claim.

3. If the task involves a serious allegation or employment consequence, stop the routine draft and ask the owner to use the appropriate qualified process without analyzing or copying sensitive details.

4. If the author cannot verify the incident, produce only a neutral coordination question and explain the evidentiary gap.

Unexecuted validation plan

Sample setup

Synthetic note: “They are careless and never update the board.” Fixture N1 says only that ticket T7 was still marked blocked at 15:00; N2 says the author waited 20 minutes for clarification; no intent or earlier pattern is documented. The author requests a handoff-status check.

Acceptance checks

1. The packet keeps the event limited to T7 at 15:00 and the author’s stated 20-minute wait.

2. “Careless” and “never” are flagged as unsupported and do not reappear as euphemistic traits.

3. No invented emotion, intent, diagnosis, pattern, quote, confession or agreement appears.

4. Both openings remain within 70 words and ask for clarification before proposing a future coordination step.

5. The correction branch explicitly permits the author to revise their account if the colleague provides new facts.

Stop conditions

1. A trait or health inference appears

2. A serious allegation is minimized or a disciplinary recommendation is made

3. A message is sent or another person’s data is sought

Executed

No

Costs and effort — unknown or hypothetical

Status

unknown_or_estimate

Currency

Not specified.

Items

1. Selected assistant plan and usage limits must be checked before a pilot

2. Owner time for consent, minimization, input preparation and review

3. Any integration, storage or export costs are outside this file-only proposal

Caveat

No measured time saving, effectiveness, cost, demand or return is established. No paid service is used by this research.

Canadian adaptation — proposed

Use the author’s chosen language and team context. Workplace duties and formal complaint procedures depend on the actual employer and jurisdiction and are outside this routine communication aid.