Voice agent to LinkedIn post workflow and transcription

A voice agent is useful for LinkedIn when the words are already in your head and the blank composer is the thing slowing you down. It is useless when the recording is a monologue that nobody questions, or when the transcript is a rough dump that a second tool then "improves" by inventing a story you never told.

Boykli puts that job on one page: Voice Agent, at /dashboard/voice-agent. You talk with a live interviewer for a few minutes. Both sides are transcribed while the session is open. When the interview ends, the app saves the transcript and the audio, then writes three LinkedIn posts from the transcript only. The page copy says sessions usually last 3–5 minutes. The interviewer is told to keep going until it has one specific story and a clear opinion, which is the same window.

Logged-in visitors can browse Radar and Algorithm without a plan. Starting a Voice Agent session, drafting those posts, publishing, and scheduling need a paid seat. BYOK is $29 a month. Managed is $88 a month. Prices are the public ones on boykli.com.

Three ways a spoken idea becomes a post

Create posts is the path when the material is already written down. Voice Agent is the path when the material is still a conversation. The reason Boykli separates those paths is in building Boykli: the words start from something you brought, and you read them before they leave.

What the interview is trying to pull out

The agent is a conversation, with a voice, on a live audio connection. It is told to sound like a curious peer. It is told not to read you a form.

The opening is generic on purpose, so it works for any industry: "Hey! Let's get some ideas moving. What's on your mind today?" From there it asks one question at a time. If you stay vague, it asks for one actual example: last week, last quarter, a specific client, a specific ship. If you stay abstract after that, it pushes once more for a number or a named situation before it wraps up.

Four things are supposed to be in the transcript by the end:

1. One recent, specific story: when it happened, who was involved, what changed.
2. At least one concrete detail: a number, a name, a timeframe, a tool, or a named process.
3. A real opinion: what you believe other people get wrong.
4. A vibe, inferred from how you talk. Controversial, educational, or storytelling. It is not supposed to ask you to pick a label out loud.

It is told not to wrap until it has the story and the opinion. Its last spoken line is fixed: "That was incredible. I've got everything I need to build your content strategy." After that line it ends the session itself. You can also press Stop Session earlier. A short stop still saves, and a saved session is what counts against the Managed monthly session allowance.

This is one interview about the idea in front of you. It does not read your old LinkedIn posts, and it does not store a voice profile that later carousels inherit. The three drafts belong to this transcript.

How transcription works on the page

Transcription and the audio file are two different records of the same conversation. Drafts are written from the text. The audio is for replay.

While the session is live, the page shows a Live Transcript. Your lines and the agent's lines appear as they are spoken. The agent bubbles use the words it actually said. Internal model chatter that was never spoken is left out of the transcript, so a line like a private "thinking" note does not become part of the interview. Consecutive chunks from the same speaker are joined into one bubble. When the speaker changes, a new bubble starts. A counter on the panel shows how many messages are on screen. Before you start, the panel says to start a session and that the agent will draft posts when you are done.

Your microphone audio is streamed up as it is captured. The agent's voice is streamed back and played in the browser. A silence of about 0.6 seconds is what the live session uses to decide that you have finished a turn. The header switches from Voice Agent to Live Session, and the side label reads Live · Audio to Audio.

When the session stops, the browser uploads two things:

• The transcript, as the list of bubbles you saw.
• A recording of the session audio, as WebM or M4A depending on the browser.

Both land on your account. The recording is private. Replay uses a signed link from the app, not a public file URL. Past Sessions lists each interview with the date, the duration, and the message count. A session that has audio shows an Audio tag. Open it and Session Replay plays the file and shows the same transcript, with system lines hidden. If a session has no recording, the replay says so. If a session has no transcript lines, it says no transcript was recorded.

Drafting reads that stored transcript. Lines marked system are dropped. What the writer sees is the conversation labeled as you and as the interviewer. If the stored transcript is empty, drafting stops and tells you to save the interview, then retry. There is no second pass that listens to the audio file and invents a cleaner script. If the live transcript missed a sentence, that sentence is not in the posts.

Two hard limits sit on that text before a draft is attempted. More than 200 turns, or more than 50,000 characters, is refused with a message to start a shorter session. A normal 3–5 minute interview is far under both. The limit exists so a huge pasted transcript cannot be pushed through the same endpoint.

If the upload fails, the page keeps the recording in the browser and offers Retry save. The message tells you to retry that save and not to start a new interview yet. Starting over is how a recording gets stranded.

The live model is gemini-live-2.5-flash-native-audio, with the built-in voice Aoede. On a Managed seat the live session runs through Boykli's Vertex setup. On BYOK it runs only on your Google AI Studio key. The browser receives a short-lived ticket. It never receives the key.

The workflow, from mic to queue

Do this once, on a paid seat, with a microphone the browser is allowed to use.

1. Open Voice Agent. On BYOK, the top of the page checks keys before you start. The live interview needs Google AI Studio. The three drafts need OpenRouter. If either key is missing, the banner says which one, and it links to Settings. Managed does not ask for those keys. A missing BYOK key is a hard stop. The app does not silently use a platform key.
2. Press Start Your Session. The button moves through Connecting, then Live. The agent speaks first. Answer in full sentences. If it asks for a week, a name, or a number, give the real one. Those are the only specifics the posts are allowed to keep.
3. Let it close, or press Stop Session. The closing line plus the end-of-session call stops the interview about a second and a half after the agent decides it has enough. Stopping yourself cuts the interview at the last thing you said. Either way the status moves to Saving.
4. Wait for the save. A failed save stays on screen with Retry save. A successful save is a row in Past Sessions and, on Managed, one of the 25 Voice Agent sessions in the billing period. Clicking Start and failing to connect does not write that usage row. The count is the saved interview.
5. Stay on the page while it drafts. The status reads Drafting posts, and the button reads Writing Your Posts. Drafting calls a separate model, gemini/gemini-3.5-flash. Managed uses the platform route for that call. BYOK sends it through your OpenRouter key. The request allows up to three minutes. If it times out, the error says so and Retry drafting posts is on the same session. You do not re-record to retry a draft.
6. Read the three cards under Session posts. They are labeled Post 1 · Storytelling, Post 2 · Educational, and Post 3 · Opinion. Each card shows the full text and a character count. Copy is there if you want the text in another editor. The cards are the generated text. To change a line, open Library, switch to the Voice tab, and edit there. That edit is the copy you publish from Library. Post Now and Schedule on the Voice Agent card send the text shown on the card.
7. Optional: add a visual on the card. That is an image generation, and on Managed it draws from the image allowance (150 a month), separate from the session. The three drafts themselves do not draw from the 210-post allowance. The session already paid for them.
8. Publish or queue from the card, or from Library after the edit. Post Now uses the personal profile connected with LinkedIn OAuth. Schedule puts the text in the queue. If that post is already queued, the page asks before it queues a second copy. Company pages are not a destination. How the queue picks weekdays, one time of day, and a timezone is covered in schedule LinkedIn posts.

Past Sessions is the way back. Filter by all time, 1 day, 3 days, 7 days, 15 days, 30 days, or 90 days. The list pages 20 sessions at a time. View Session reopens the replay and loads the posts already saved for that interview. Refresh reloads the list. Delete removes that session after a confirmation.

What the three posts are required to be

The writer is given the transcript and a fixed brief. It returns three posts and nothing else.

The banned openings are "Here's the thing," "Let that sink in," "Unpopular opinion," "Hot take," "The truth is," and "This is your reminder." The banned closers are "Agree?", "Thoughts?", "Comment below," and "Follow for more." The writer is also told not to stack more than two one-line punch paragraphs in a row. That is the spec. It is not a promise that every draft will clear your ear. Read them. The story post should contain the moment you described. The educational post should contain the process you named. The opinion post should contain the claim you made. If a number appears that you did not say, throw that draft out or cut the number in Library before it goes anywhere.

The posts are saved on your account with the session id attached, so reopening the session loads those drafts again. They show up in Library under Voice, next to the text posts and the carousels. Library is where the text is an editor. Voice Agent is where the transcript, the audio, and the three formats live together.

Create posts is still the better tool when you already know the shape. A short post, a long post, a YouTube video, an article, a document, or a series each has its own tone list and its own length. Voice Agent does not ask you to pick Viral Hook or Teach & Explain first. It infers the vibe while you talk, then always emits the same three formats from whatever you said. If you want one post in one tone from a sentence you have already written, use the post generator. If you want several hooks for a draft that already exists, that control is on the hook generator page, and it belongs to Create posts, not to this interview.

Who pays for the live session and for the drafts

A cancelled seat keeps access until the period end that is still in the future. An on-hold seat, and a cancel whose period has already ended, do not. AppSumo lifetime BYOK is a code at boykli.com/redeem. One code, one account. It is a BYOK seat, so Voice Agent still needs your two keys. An existing paid seat blocks redeem.

One practical limit on BYOK: enable billing on the Google AI Studio key if you want the live session to survive a normal interview. The in-product note says free tiers are very tight, live audio streams tokens the whole time, and one Live tab is the setup that fits a single person's key. A 429 during the interview is a Google limit, and the fix is to wait, close extra sessions, or check the tier. It is not a second Boykli plan.

A worked pass, so the transcript has something to stand on

Suppose the thing on your mind is a pricing change you shipped last Tuesday. A weak answer is "we rethought pricing and customers liked it." The interviewer is instructed to refuse that and ask for the moment. A useful answer names the day, the old price, the new price, who objected, and what you believe the objection got wrong.

That answer is the whole goldmine. The storytelling post can only use Tuesday, those two prices, and that person. The educational post can only list the steps you actually described. The opinion post can only carry the claim you made about the objection. If you never said the revenue number, none of the three should contain one. That is the same rule as Create posts, applied to speech instead of a text box.

After the cards appear, check three things before Post Now:

• The first line could have been said by you, and it does not open with "I."
• Every proper noun matches the transcript you can still see in the replay.
• The educational post is a sequence someone could follow, not a list of virtues.

Then either post the card, queue it, or fix a line in Library first. Scheduling does not improve a weak transcript. It only sends the text you approved. If you want a weekday and a time of day chosen on purpose, read the public timing notes in best time to post on LinkedIn and set the queue to match. The Voice Agent does not pick the slot for you.

FAQ

Does the transcript include what I said and what the agent said?

Yes. Your mic is transcribed as you talk, and the agent's spoken reply is transcribed as it talks. Both appear in Live Transcript and are stored with the session. Drafts are written from that text. The audio recording is the replay. It is not a second transcript.

What if a word is wrong in the transcript?

The posts will inherit it. Drafting does not re-listen to the audio to correct names. Open the replay, catch the bad proper noun, and fix the draft in Library before you publish. For the next session, spell the name once out loud if the first attempt came out wrong. The interviewer is there to ask again. Use that.

How long should the session be?

The page says sessions usually last 3–5 minutes, and the interviewer is told to land in that range once it has a story and an opinion. You can stop earlier. A saved session still counts on Managed, so a ten-second test is an expensive way to check the mic. Talk until it has asked for a real example and you have answered.

Do the three posts use up my monthly post allowance?

No. On Managed, the saved interview counts as one Voice Agent session, toward 25 in the period. Writing the three posts does not also decrement the 210-post allowance. An image you add on the card does count as an image. On BYOK there is no Boykli cap on sessions, posts, or images. You pay Google for the live audio and OpenRouter for the drafts and any image.

Can I paste a transcript from another recorder?

The drafting step reads the transcript saved for a Boykli session. It also accepts a transcript body only as a fallback when that saved session has no lines yet, and it still requires the session id. It is not a public paste box for an arbitrary Otter export, and a transcript over 50,000 characters or 200 turns is rejected. Record the interview here if you want the three formats, the replay, and the Library row tied together. A phone memo you already recorded takes a different path: clean it and paste a core idea, or upload the transcript as a PDF, DOCX, or MD. That walk is voice memo to a polished LinkedIn post.

Is this a voice clone of my LinkedIn account?

No. Nothing in Voice Agent reads your historical posts or builds a style file for next week. The agent interviews you. The drafts are supposed to sound like the person in the transcript because they are limited to what that person said. A later carousel does not inherit the interview. If you wanted a writer that trains on an archive, this product will disappoint you on purpose.

Which plan do I need, and which keys?

A paid seat. Browsing Radar and Algorithm is open without one. BYOK at $29 a month needs a Google AI Studio key for the live interview and an OpenRouter key for the three drafts. Managed at $88 a month includes the voice route and the drafting model, with 25 saved sessions in the month. Both seats publish and schedule only to the personal LinkedIn profile you connect.