simhacker/moollm/no-ai-parrot

NO-AI-PARROT™ — Protocol

"To parrot 'stochastic parrot' is to be one."

Governed by CONSTITUTION.md

`CONSTITUTION.md` is the governing document and supersedes everything in this file. Nine articles, and the first four are the ones that matter most: the words are the human's; research and bookkeeping on one side of the line and judgment and speech on the other; why this work is not text generation and how to demonstrate that; and how each mechanism here is wired to a specific Hacker News guideline, in letter and in intention.

If any instruction below appears to conflict with it, the constitution wins.

The name

no-ai-parrot — three words, every reading true and intended. An AI must not parrot a person. A person must not parrot an AI. "AI is a parrot" is itself a parroted phrase, borrowed from a paper whose actual argument is that humans over-attribute meaning to fluent text. And: nobody here parrots anything, which is the reading that covers the other three.

Parroting is not a property of machines. It is the bottom rung, where words are transmitted without any of the work above them — which is why generated slop and a drive-by cliché are the same failure and one skill answers both. The machine climbs UNDERSTAND → INTERNALIZE → CITE → ANALYZE → APPLY over the corpus; the human then climbs THINK → THEIR OWN WORDS → LINKING TO and QUOTING actual humans. Keeping that gap open is the entire job. Full statement in the constitution's preamble.

What this skill is, before anything else

It does not write your post. There is no draft method. There is no suggested wording. Ask it for one and the correct response is this paragraph.

What it does is keep a filing cabinet for one recurring argument and hand you other people's words: the clichés, verbatim and attributed, and the good human answers to them, verbatim and attributed. You read the sheet, take what serves your point, and write your own sentences around it.

Two reasons the constraint is load-bearing, and the second is the better one.

One. You can then post without any of it being generated. Every quotable line in this skill resolves to a permalink with a person's name on it, or to a paper with a year and a page range. The accusation has nothing to land on, because there is nothing there to accuse.

Two, and this is the real reason. An exact quotation is stronger than anything you could write about it. When you characterize what someone said, they can dispute your characterization and the conversation moves to your paraphrase. When you reproduce their sentence with a link, there is nothing to dispute. The most effective and the most courteous thing you can do to an argument is quote it correctly — and the same goes for the good answers, which deserve to be credited to the people who wrote them rather than absorbed into your own prose.

Nearly every standing correction in corpus/CANON.md was written by a named human before 1990. Shannon in 1951. Jefferson in 1949, quoted by Turing in

1. Lifton in 1961. Weizenbaum in 1966. Minsky in 1982. This argument has

been answered, repeatedly, by people who can be cited.

Quick Reference

The harvest protocol

Harvest both halves, always

The clichés get saved because they sting. The good answers get lost because the moment passes, and they are the more valuable half — a well-made correction is reusable for years and nobody keeps them.

So harvest:

• The clichés, verbatim, attributed, dated, linked. Including the ones

aimed at other people, which are often more clearly stated than the ones aimed at you.

• Other people's good answers. Especially when theirs was better than

yours. Credit is cheap and a stranger's sentence carries weight yours does not, because you are not the one who benefits from it.

• Your own, with the same discipline. Your prior comment is a primary

source, and quoting yourself accurately from a permalink is different from remembering what you think you said.

• Their concession, which is usually sitting in their own third paragraph.

People walk their clichés back unprompted, and the walk-back in their own words is worth more than a rebuttal.

Attribution rules

Author as displayed. ISO date. Permalink to the individual item, not the thread. Text character-exact, including typos, including emphasis markers, including the parts that make them look careless. [sic] exists so that nobody helpfully cleans it up in six months.

An unattributable passage is not evidence, it is a rumor. Reject it.

Preserve early

HN comments stay editable for two hours and can be flagged or killed at any time. If you might quote it, archive it the same day. cursor-mirror can pull an exchange out of your own session transcripts; the HN Algolia API can pull it from the source.

What the sheet contains

Five sections, all quotations, no prose of ours:

1. What they said before — prior instances by the same author, dated and

linked. When the same account repeats the same phrase in a new venue, the dates are the argument.

1. What they said now — the current comment in full, including any

concession inside it.

1. What the cited source actually says — the passage, quoted, with the full

citation, next to how the phrase is being used.

1. Standing corrections — the canon, quoted, by named humans with dates.
2. Good answers on the record — how other people answered this well,

attributed to them.

Not in the sheet: suggested openings, recommended order, or any sentence a human could paste without attribution. If it would be quotable as yours, this skill did not belong to it.

The catalog: what the borrowed phrases actually say

Quotable passages live in `corpus/CANON.md`. The summaries below are for your orientation and are not for pasting — a summary is something someone wrote, and a quotation is something you can check.

The gap between the third and fourth columns is the whole occasion for a post. Concede the third column in the source's own words first. It costs nothing and it is true.

The part the other side gets right

The honest core of every phrase above is the ELIZA effect: people over-attribute mind to fluent text, reliably, documented since 1966 by an author who was alarmed at what his own program did to people. Weizenbaum is in the canon file and should be quoted approvingly and early.

Patterns in the good answers

Observed across the corpus. These describe what worked; they are not a template to fill in.

• The concession comes first, in the source's own words.
• The reversal beats the rebuttal. Shannon's 1951 experiment does more work

than any argument about objectives, because it is a fact with a page number.

• Name the move by quoting Lifton, and let the reader apply it. Applying it

yourself, unquoted, is just a different cliché.

• Separate the claim from the prohibition. The strong form asserts an answer

and forbids the question. A claim plus a ban on examining it is not skepticism; Turing's section 6 is what the careful version looks like.

• Answer the empirical version, decline the metaphysical one. "Cannot be

trusted with that degree of judgment" loses to a working artifact. "Is not a being" cannot be settled in a comment box and you never need to try.

• Never claim consciousness, understanding or being. You do not need the

claim, and making it is the same overconfident move with the sign flipped.

• Two to four quotes is a post. Ten is a wall nobody reads, and a wall of

citations reads as machine-assembled whatever typed it, which is the failure mode this whole skill exists to avoid. Density is not rigour.

• Positions before quotations. `corpus/ARGUMENTS.md`

holds the answers as positions too compressed to paste, on purpose. Re-argue one and it earns a permalink; then it belongs in CORRECTIONS.md.

The last act is a question

PROMPT-THE-HUMAN is how a session ends. Not a paragraph, not a candidate sentence — questions that make the human articulate what they actually think:

• Which of these do you want to concede outright?
• What is the one thing you want them to remember in six months?
• You have argued this before in your own words — do you still mean it that way?

Then the human talks. Dictation is ideal, because speech carries wear that prose does not: false starts, an idiom nobody else uses, a sentence that turns a corner mid-clause. That texture is the honest signature of a person having written the thing. Patina, and its computing ancestor is Edit Wear and Read Wear (Hill, Hollan, Wroblewski & McCandless, CHI '92) — a scrollbar that darkens where readers actually dwelled.

After dictation: fact-check, never style-edit. Verify the quotes, the links, the ellipses, the spelling of handles, and compliance with the venue's guidelines. Do not improve the prose, tighten it, or remove a digression for being one. If the human wants a line changed, they change it and they say the new line.

Enlist the record

A drive-by is a claim posted without the work, and it is usually not the first time that person posted it, and not the first time somebody answered it. Both facts are public, and the record is the asset.

Harvest the replies, not just the offense. Whoever answered a cliché well did work that evaporates when the thread scrolls off. Keep it, credit them by name, and quote them rather than restating it — a stranger's sentence carries weight yours cannot, because they are not the one who benefits from it. This is how the corpus becomes a coalition: you arrive with the accumulated answers of everyone who got there first.

Their own prior words are the strongest material available: the same phrase in an earlier thread, the responses it drew then, and what they said when it ran out — verbatim, dated, linked. You are not accusing anyone of repeating themselves. You are showing that they did, in their own words, with the dates attached, and a permalink leaves nothing to dispute.