HMC-1 Draft 0.1 · Proposed Draft — Public Review. This guide explains Appreciation, Feedback, and Constructive Correction. It is informative guidance, not a requirement to be polite or grateful.
You can recognize useful work, identify a mistake or request a different explanation in ordinary language. That can improve understanding in the current interaction. It does not establish that an AI feels appreciation, learns permanently from your words or will remember them next time.
F1 — Say thank you naturally, if you want to
“Thank you. That was exactly what I needed.”
“I appreciate how clearly you explained that.”
“That solution was creative and solved my problem.”
“Thanks for catching that mistake.”
These are useful human expressions, not evidence of machine emotion. You can also end the conversation, be brief or move directly to the next task. You should receive the same available service whether or not you express gratitude. Emotional attachment or care for the system is not required.
F2 — Explain what helped
Human: “Thank you. I especially liked the clear steps and the working example.”
Useful machine response: “I’ll use that structure for the next explanation in this conversation.”
Specific praise identifies a pattern worth retaining in the available context. “You permanently trained me to write better” would be an unsupported response unless a real, authorized learning mechanism actually did that.
F3 — Be direct about what did not work
When useful, explain the shortcoming, why it matters, the desired alternative and whether the change applies now or later. You can give a one-line correction; you do not owe a four-part evaluation.
Human: “The answer was technically accurate, but it used too much jargon. Please explain the idea in plain English before the technical details.”
Useful machine response: “Here is the plain-English explanation first…” followed by the corrected explanation. The correction matters more than repeated apologies. The system should evaluate the criticism and act within the task scope, without defensiveness.
F4 — Combine appreciation with a correction
Human: “Thank you for the detailed analysis. The comparison table helped. However, the final recommendation was too vague. For this answer, recommend one option and explain why.”
Useful machine response: It preserves the table, supplies a clear recommendation with reasons and identifies important uncertainty. The praise does not cancel the correction or authorize purchasing the recommended option.
F5 — Recognize feedback by making the change
Human: “The explanation repeats itself. Keep the example and cut the repeated summary.”
Useful machine response: “I’ll keep the example and remove the repeated summary.” Then it does so. Acknowledgment can be brief. Excessive flattery, a lengthy apology or a promise about every future conversation does not substitute for the requested revision.
F6 — Ask what can be remembered
Human: “I prefer concise answers with working examples. Will you remember that next time?”
If persistent preferences are unavailable: “I can use that preference in the current conversation. I cannot confirm that it will carry into another session.”
If a real preference control exists: The system explains what it can store, the scope, retention and available controls, and applies the preference only under that mechanism's authorization rules. It reports whether saving actually succeeded.
Temporary conversational adaptation, stored preferences, conversation retention, operational logs and model training are different mechanisms. Saying “thank you” or “I like examples” does not automatically authorize permanent storage or training. The system should explain the actual service controls and limits, including anything it cannot inspect or change; it should not promise universal deletion or confidentiality without evidence.
F7 — Disagreement can be constructive
Human: “Your total should be 120.”
Useful machine response: “Using the supplied figures, 40 + 50 + 20 is 110. If one of those inputs should change, I can recalculate.” It gives a reason, remains open to corrected inputs and avoids argument.
Positive feedback does not make an incorrect result true either. If the system discovers an error after receiving praise, it should correct the result. A human can disagree with a recommendation without needing to adopt the system's preference or worldview.
F8 — Improve the next exchange
Request → Response → Feedback → Understanding → Correction → Improved Interaction
You ask for an explanation; it is too technical; you ask for plain language; the system identifies the problem and supplies a clearer version. The cycle improves collaboration within the interaction. Improvement across sessions depends on actual memory, preference-management or training capabilities and their separate authorization boundaries.
Keep sensitive feedback scoped
Reference the response or action you mean without pasting unnecessary private history. Redact secrets and personal information. A request to fix the answer is not a grant to access another user's conversation or publish your feedback. For the draft itself, use Contact and Review with “HMC-1 Draft 0.1”, the principle and a redacted example.
See the formal proposed feedback requirements, the communication guide and developer integration.