SW Agent's human thinking mode

With human thinking on, SW Agent answers the way a careful person writes. It reads what you asked: the kind of ask, what it is about, and the length and reader if you said them. It takes the closest model answers it has and plans your answer from the best one's shape: its parts in order, what each part must answer, and how it opens and ends. It never copies a model answer's words, only its shape.

Then it reads our docs, its world knowledge and every knowledge pack for each part, all at once, and writes each part from what it read: the answer first, then what adds to it, every sentence with the number of its source. Last, it checks every sentence against its source before you see it.

• The thinking steps under the reply show the plan: one step per part, with the questions that part answers and what was found for it.
• How I thought it through, under the reply, names the model answer it planned from and why it chose it.
• The whole answer comes as a document to download, with its sources and what it could not find.

Human thinking is in preview, so an answer can sometimes stray off your subject. Press Human thinking beside Auto and Manual, under the message box, on SW Agent's page, in the SW Agent tab in Classic and Studio, or in a tool page's SW Agent. It stays on in this browser until you press it again.

It answers asks for an explanation, how-to steps, a plan, research with sources and writing. A fact, a sum, a chart, titles, names, a logo and jobs with the site's tools are answered as usual even when it is on, and a blog post from your own attached documents is written from them as usual.

With it off, an answer that would gain from it offers “Think it through, part by part” under the reply: press it to answer that one question this way. With it on, “Answer it the usual way instead” does the opposite once.

• Open SW Agent

Each model answer is filed under a kind of ask and filtered by subject area, length, reader and form, with one to three in each, chosen for quality over count. Every sentence of a model answer names the passages it came from, all of them openly licensed records SW Agent ships or our own help, and a test proves each sentence's words, numbers and names are in them.

When no model answer of your kind of ask exists yet, the closest kind's is used, and the thinking steps say so: a plan is planned like a how-to, research is explained like an explanation.

• How-to steps: 2 model answers (medium guide), such as “How do I record my voice or an instrument with a microphone?”
• An explanation: 2 model answers (medium answer), such as “How does digital audio sampling turn sound into data?”
• Writing: 1 model answer (medium blog post), such as “Write a blog post about the history of electronic music.”

Every sentence you read is a source's own words, with that source's number, or one of a few fixed sentences SW Agent says about its own work. A sentence whose number, date or name is not in its source is cut, never reworded. A source cited beside another that does not give the same number is taken off that sentence. Two sources that give different numbers for the same thing are both shown, each with its number, never averaged.

A part it found nothing reliable for says so plainly, and the answer ends with what to try next. A sentence quoted without the paragraph it needs ("this method", "the following") is never used.

• “I found nothing reliable on this part in my sources.”
• “Ask about it on its own, with the full name of what you mean, or search the help centre.”
• “My sources disagree on this, so both are shown.”
• “Every sentence carries the number of the source it came from.”
• “Here is what my sources say, in fewer words, with every source I read listed.”

A rating of an answer from human thinking also says what kind of question it was read as and which model answers it was planned from. SW Agent adds the ratings up for each kind of question, each part of SW Agent and each model answer, over the last 30 days: one that fewer than 70% of at least 20 different members liked is under the line.

A model answer under the line is no longer planned from while another of the same kind, form and length is not, so a kind of question never loses its last model answer to ratings alone. A kind of question under the line is answered in a safer shape: it opens with what the sources say, keeps every part to 2 sentences and lists every source it read, cited or not. Its disliked answers go to the top of the list the team works through, and the team can switch the safer shape on or off for each kind.

Each member counts once, by their latest rating, however often they rate, and a rating sent while signed out is counted but never decides anything. Ratings never change the words of an answer or the benchmarks SW Agent must pass: the team turns disliked answers into tests and fixes, and the benchmarks prove each fix.

• Your ratings, in your SW Agent settings
• What a rating holds (Privacy Policy)

The model answers cover how-to steps, an explanation and writing for now; comparisons, troubleshooting and research summaries of their own, and posts of other lengths, come next. A part comes back thin when SW Agent's sources do not cover it: the answer says so instead of filling it.

A sentence is kept only when it is about what you asked, not merely sharing a word with it: a game engine's audio recording never answers a question on recording vocals, and our tips on visualizer looks never fill a post on the history of hip hop. So a part can come back thin where a sentence about a neighbouring subject used to fill it. In this preview a word with two meanings (a gas compressor for an audio compressor) can still slip through, as its source's own words.

Rate its answers with the thumbs under each one, as any other answer; How your ratings act, above, says what they change.

• How SW Agent answers
• SW Agent's benchmarks