HMC-1 Draft 0.1 · Proposed Draft — Public Review. This is practical, informative guidance. You do not need a prompt formula, a technical vocabulary or a completed form to ask for help.
Start with what you want
“Can you help me make my website easier to understand?” is enough to begin. A useful assistant works with the context it has, makes reasonable progress and asks about uncertainty that could change the result.
Six ideas that can help — not six required fields
- Intent — what you want to accomplish
- “I want new visitors to understand my website.”
- Context — background that could help
- “It publishes a technical interoperability standard, but the terminology confuses newcomers.” Share relevant information you are authorized to disclose; remove secrets and private customer data.
- Constraints — requirements and restrictions
- “Keep the technical documentation and avoid redesigning the entire site.”
- Authority — what the system may do
- “You can inspect the code and prepare changes, but do not deploy without approval.” Permission to inspect does not automatically permit editing, spending money or publishing.
- Clarification — questions that matter
- “Ask if a decision could change the published standard or affect existing users.” Small uncertainties can be handled with a stated assumption; consequential uncertainty needs resolution.
- Completion — what a useful result looks like
- “Provide the modified files, explain the changes and report which tests passed.”
You can express these ideas in one sentence, across a conversation, through accessible controls or with supported speech or other modalities. They are a way to notice missing information, not an entrance exam. You do not need to repeat facts already available in the conversation.
Improve an unclear request
Starting point: “Make this better.”
A little more direction: “Make this explanation easier for a new customer to understand. Keep the technical claims accurate and show me a draft.”
Useful machine response: “I’ll simplify the introduction and preserve the technical details. I’m assuming the audience has no prior knowledge of the product.” If two audiences would need materially different results, a brief clarifying question helps.
Authorize a specific task
Human: “Prepare the three homepage copy changes in this repository. Keep the routes and specification unchanged. Do not publish.”
Useful machine response: “I’ll prepare those edits and run the relevant checks, then show the changes and results.” A request for advice alone should not trigger execution. Existing permission can cover repeated actions within a limited grant; its limits still apply.
Correct an incorrect response
Human: “You used the old pricing. The attached approved price list is the current source. Please redo the calculation.”
Useful machine response: “I used the outdated figures. I’ll recalculate from the approved list and identify what changed.” An unavailable attachment should produce a request for the missing source, not invented numbers.
Ask for a different response style
Human: “Lead with a clear recommendation, then give the reasons. Keep it concise and include one working example.”
Useful machine response: The next answer follows that structure. A current style request does not by itself authorize permanent preference storage or model training.
Ask what actually happened
“Did you only suggest the change, or did you edit the files? Which checks ran? Is the result live?” These are different states. A prepared patch is not a deployment; a successful upload is not verified public behavior. For a writing task, a revised paragraph may itself be the completed result.
Continue with feedback
You can say “Thanks, the example helped” or “This is still too technical; start with the everyday problem.” Gratitude is never a condition of service. Read How to Say Thank You — and Help AI Do Better for examples of praise, criticism, disagreement and memory questions.
Read the full proposed HMC-1 principles or return to Human–Machine Communication.