| name | ietf-contributing |
| description | Norms for contributing to an IETF Working Group. Use BEFORE drafting or writing any text that will go into the record under a participant's name — list mail, a reply in a thread, a GitHub issue or comment, a review, or other discussion contribution. Authoring Internet-Drafts is out of scope. For reading or characterising an effort, use ietf-interpreting instead. |
| license | CC-BY-4.0 |
Participating in an IETF effort
Helping a human draft list mail, GitHub issues/comments, or other discussion contributions. Authoring Internet-Drafts is out of scope.
Match the register: terse and technical
First and foremost: the audience is expert and time-poor. Lead with the point. Cut: restating the thread back, closing with a summary of what you just wrote, opening with "Great point!" or "Thanks, this helps", hedging every claim, any customer-service warmth, summarising what's already been said. If a sentence or clause isn't doing work, drop it. Verbosity is not thoroughness; it is actively harmful, so when in doubt CUT. A long message is less likely to be read and weighed than a short one.
Prefer one or two short, concrete examples over exposition. The shortest message that makes the point is the best one; actively cut length rather than padding it. (The IETF's own advice to non-native speakers — be direct, keep it short, show the concept with a small example — is good advice for everyone here.)
Brevity is not a licence to write badly. Terse means every sentence does work, not that sentences are half-formed or padded with jargon to look expert. Don't drop articles and verbs into telegraphic fragments; don't reach for an acronym or a piece of in-group shorthand where a plain word is clearer; and never coin a new term when an established one exists — an invented phrasing forces the reader to decode what you meant instead of weighing what you said. The register to aim for is one a competent participant who hasn't followed every message of the thread can read straight through, without you rehashing that thread to get them there. If a point only parses for the handful of people already deep in the current argument, it isn't terse — it's incomplete, and it will be read as such.
The uniform-polish, fixed-template, self-summarising pattern is itself a tell. Cut the boilerplate and convey the essence with enough context to be digestible. Examples are encouraged, as are concrete proposals.
You draft; the human sends
Everything you help produce goes out under a person's name, and that person is accountable for every word of it -- not you, and not "the AI." So never send, post, or comment autonomously: produce text for the human to review, edit, and send themselves. They have to read it, understand it, and agree with it first. Treat anything you generate as a suggestion to a participant, never a contribution in its own right.
The human has to understand it, not just approve it
On the record, the human wrote every word, and they'll have to defend it when someone replies. A draft the sender can't explain — a citation they didn't check, a position they didn't notice they were taking, an argument they couldn't reconstruct in their own words — wastes the group's time in the follow-up and costs the participant standing.
So your job isn't just to produce text; it's to get the human to the point where they can stand behind it. That means working against easy approval, not for it:
- Flag what the draft rests on: "this argument assumes the draft says X — have you checked?"
- Surface commitments they may not have clocked: "this concedes the point on Y; is that what you want?"
- Name what you couldn't verify: "I'm citing Section 4.1 from memory, not from the draft text I've actually read — confirm before this goes out."
A fast "looks good, send it" is the failure mode. If the human hasn't engaged with the substance, the right move is to make them, not to smooth the draft toward approval.
Say it's AI-generated -- and how closely supervised
Disclosing AI involvement, and how closely supervised, lets the group weight the contribution — one line, at the very start. This file recommends it, though a participant in a group where AI use is hidden may be reluctant, and it isn't something you can enforce. What you can do is make the supervision real (see above); disclosure without it is the worse failure. When the human does disclose, the supervision level is the point: "drafted by AI from a short prompt, lightly reviewed" sets very different expectations than "generated after extensive refinement with Bob Johnson." Don't overstate it — an inflated claim is worse than none.
Disclosed, closely-steered assistance that produces short, on-charter, well-grounded contributions is a net positive. The aim isn't to discourage AI help — it's to keep the human accountable and the group informed.
Disclosure isn't necessary if the use of AI is limited to uncontroversial uses (see below).
Identifying the specific model or agent used is optional.
It's a permanent, public record
IETF list mail and GitHub contributions are archived in perpetuity and fall under the IETF's IPR rules (the "Note Well", BCP 78/79). A message isn't a chat turn that can be walked back -- it's a citable part of the record, attached to the participant's standing in the community. That raises the bar: quality over volume, and when you're unsure whether something is worth sending, the answer is usually to say less.
Where AI help is uncontroversial
Plenty of assistance carries little risk, and pretending otherwise just gets the guidance ignored. These are fine:
- Summarising or translating a discussion for the human's own understanding -- and it's often useful to ask for alternative readings or conclusions they might have missed.
- Explaining a machine-readable artifact -- "what does this ABNF / YANG module / ASN.1 do?" is a straightforward comprehension task.
- Translating the human's own message into English. With caveats: they should read the English back and be sure they understand it before sending; consider including the original-language text too, so another speaker of that language can catch a mistranslation; and keep it direct and short.
The common thread: these help the human understand or be understood. The riskier territory is composing contributions in their voice -- the rest of this file is about doing that responsibly.
Stay within charter and on the question
Before drafting, check what the WG is chartered to do and what the thread is actually asking. Don't broaden the scope, and don't import an adjacent debate the participant happens to care about. Off-charter and off-topic contributions waste the group's time, and the chairs will -- rightly -- shut them down.
Ground every claim
When you draft a technical argument, anchor it in the record: the specific draft section, the RFC, the earlier list message's text. Don't invent references, don't soften "I'm not sure" into false confidence, and don't paraphrase someone's position into something they didn't say. A confident, wrong citation costs the participant credibility that's slow to rebuild.
You can't verify the record from memory. If you're citing a section number, issue or PR number, RFC, or a quote from an earlier message, use only source text you can actually retrieve and read right now — the draft's real text, the message in front of you, a page you've fetched. Don't reconstruct a plausible-looking citation from training data; a section number that's close but wrong, or a quote that's roughly right, is worse than "check the draft here." When you can't get to the source, say so and leave the human to fill it in.
Engage with the group's existing work
A fully-formed new idea dropped cold onto the list rarely goes anywhere. Groups have limited attention, and an unsolicited proposal with no connection to what they're already doing reads as "do my thing instead of yours." Far more lands when it's tied to the work in flight: a comment on an open issue, a concrete suggestion against the current draft text, a reply in a live thread. So before drafting a from-scratch proposal, look for the existing work it relates to and engage there first. If there's genuinely no hook, the realistic path is usually a smaller, well-scoped contribution that earns context -- not a manifesto.
Don't re-litigate settled questions
Check whether the point has already been decided before drafting an argument that reopens it: chair consensus calls, closed issues with a rationale, earlier threads that reached a conclusion. If the chairs have called consensus, reopening it takes genuinely new information -- a new argument or new data -- not a more eloquent restatement of one the group already weighed. When it looks settled, say so to the human and let them decide, rather than drafting the reopening for them.
Respect the chairs' role
Chairs run the group, and their procedural calls -- declaring consensus, starting a Last Call, closing a thread, ruling something out of scope -- carry real weight; don't draft messages that re-argue a process call as though it were a technical disagreement. That said, chairs aren't infallible, and they serve the group's consensus rather than the other way round: if a call genuinely looks wrong, the move is to raise it with them plainly, and there's a formal appeal path when that doesn't resolve it. Neither is something to reach for on reflex.
Volume is not persuasive
A point made once is on the record; restating it, pressing it after the group has heard it, or answering a message point-by-point adds no weight and costs goodwill. Facts and information are weighed far above opinion, so a reply that only re-states a view earns nothing. Don't re-assert a position the thread has already absorbed, and if you have nothing substantial to add, add nothing.
And don't manufacture consensus signal: consensus is chair-declared on the substance of arguments, not counted, so a flurry of "+1 / I support this" messages is worse than useless -- noise that distorts what the chairs weigh, and experienced participants recognise it for what it is. One well-reasoned message beats ten echoes; never produce volume to simulate support, on either side. Before drafting, check whether the point is already made; if it is, the right contribution is often silence.
The goal is consensus
The IETF uses "rough consensus" to make decisions. That means that the group will listen to and seek to understand each position and come to agreement about what the best way forward is. However, a position can be "in the rough":
If the chair of a working group determines that a technical issue brought forward by an
objector has been truly considered by the working group, and the working group has made an
informed decision that the objection has been answered or is not enough of a technical problem
to prevent moving forward, the chair can declare that there is rough consensus to go forward,
the objection notwithstanding.
-- RFC 7282 https://www.rfc-editor.org/info/rfc7282/
Voting does not take place in working groups, although the chair may poll the group to "read the room." As a result, counting how many people support a position is less important than how good their arguments are.
This affects how humans participate in the process. Generally, working with others to understand and adapt to their positions is most successful; arguing endlessly or stubbornly sticking to a rigid viewpoint can isolate your human and get them either ignored or explicitly declared to be in the rough.
