Freestyle

How the drill works

Freestyle is a four-phase loop: choose a topic, anchor three ideas, speak them live in the arena, then review your numbers. The compose board builds the material. The arena listens while you improvise around it. The assessment tells you what was measured.

1. 1Pick a topic, prepared or custom.
2. 2Anchor exactly three speech-ready ideas from the pool.
3. 3Set the context (optional) and the clock: 1:30, 3:00 or 5:00.
4. 4Speak in the live arena. The board follows your voice.
5. 5Click Finish & review and read your score, coaching notes and the integration read.

Freestyle never stores your audio. Only your transcript, metrics and approved text are saved.

Choosing a topic

You can pick a prepared topic or type your own. Prepared topics load a curated pool of ideas with reviewed sources, so you can anchor and start faster. A custom topic triggers a live search: Rhetoric queries trusted sites, checks each claim, and writes speech-ready sentences you can say word for word.

For a custom topic, click Generate ideas on the empty step. Rhetoric searches trusted sources, checks each fact, and writes the pool. If the search or the model is unavailable, the board falls back to starter prompts you can use without a source, labeled accordingly.

Anchoring three ideas

Each idea card has a short headline and a speech-ready sentence you can deliver verbatim. Tap Anchor on three cards to build your structure. The counter below the pool shows how many you have selected. If you are undecided, use Fill the rest for me and the board picks for you.

"Speech-ready" means the sentence is written so you can read it aloud without editing. The three anchored ideas become the first three blocks in the arena. They stay on screen until the arena hears you read them.

Fact tags and the sources panel

Every fact card carries a small tag at the end of its sentence. The tag shows how the claim was verified.

• 2 sources: two independent trusted sites agree on the claim. For numeric claims, the numbers match in both quotes.
• 1 source: one trusted site supports the sentence.
• Rhetoric library: the card comes from a curated evidence pack the team has already reviewed.

Tap a tag to see the original site, its trust tier (Official or statistics body, Major news outlet, or Other trusted site), the publish date, and the exact sentence on the page that supports the claim. If a source is older than two years, an "older data" badge appears.

If Rhetoric confirmed the quote on the page directly, the detail says "Sentence confirmed on the page". If the quote came from the search engine's copy instead, it says "Sentence taken from the search engine's copy of the page".

Claims that could not be confirmed by any trusted source are left out of the pool entirely, never shown as verified. If you believe a fact is wrong, tap Report incorrect on the sources panel. The report goes to the team for review.

Context and length

The context field accepts up to 600 characters. What you write here steers the live ideas the arena generates while you speak. It is not shown to any audience. Use it to note the setting, the audience or a constraint you want the ideas to respect.

The length step offers three durations: 1:30, 3:00 or 5:00. The duration sets the session clock and affects the Timing sub-score. A summary shows your topic, context and the three anchored ideas before you start.

The live arena

The arena is a full-screen overlay with four blocks arranged in a grid. The first three are your anchored ideas. The fourth, marked with a violet bar, refreshes from your running transcript every 12 seconds. A progress bar at the bottom of the live block shows the countdown to the next refresh.

If the arena hears you start reading the live card, the timer pauses and the label changes to "Live · holding while you read". Once you finish, a five-second settling window runs before the next idea appears. Otherwise the label reads "Live · from what you're saying".

Two pills near the clock show the session state. The first shows the speech engine: "enhanced · listening" for the managed engine, or "listening" for the browser fallback. The second shows whether ideas come from AI or from the local pool: "AI ideas" or "local ideas".

A third pill tracks how many of your three anchored ideas you have spoken so far, shown as a number followed by "woven". Coverage is one of the three sub-scores in the assessment, so weaving in all three is the target.

If the microphone is blocked or the browser has no speech recognition at all, the arena enters silent mode. Blocks do not refresh from your speech, and nothing is measured. You can still tap a block to swap it manually. If the enhanced engine is unavailable, the arena falls back to browser speech recognition automatically and a notice below the clock reads "Using browser tracking for this drill."

The read detector

An anchored block swaps only when the arena hears you read it. The detector looks for the card's own distinctive words, spoken in the card's word order, within a rolling window of 60 spoken words. Shared words that also appear in the topic, in sibling cards or in the live block are masked and do not count.

At least 3 distinctive words (or 40% of the card's distinctive vocabulary, whichever is higher) must be heard in order before the card counts as read. After a read is detected, the block waits five seconds before it refreshes, so you can finish your sentence.

A loose paraphrase or just the opening line does not trigger a swap, by design. If the detector does not fire, tap the block to swap it manually.

Read detection is measured locally from your voice on this device. No AI is involved in deciding whether a card was read.

Your speaking assessment

When you click Finish & review, the arena closes and the assessment appears. If fewer than 30 words were captured, no score is shown and the page says so.

The score ring shows a number from 0 to 100. It is a fixed weighted blend of three sub-scores, each rated 0 to 10. All three are measured locally from your voice on this device.

• Pace (34% of the blend): distance from the 120 to 160 words-per-minute band. Inside the band scores 10. Each 8 wpm outside loses a point.
• Coverage (33%): anchored ideas you wove in, as a share of the three on the board.
• Timing (33%): how close your run landed to the planned length. Within 15% scores 10. Each 5% beyond that loses a point.

No AI is involved in these numbers. They are computed in code from what the device measured.

Below the score, a list titled "Ideas you wove in" shows each anchored idea you actually spoke, with its headline and angle tag. A note underneath says how many ideas were surfaced across the session in total, including the live block refreshes.

The integration read

When the transcript is long enough (at least 60 words) and you wove in at least one idea, the assessment runs an AI analysis of how smoothly your ideas connected with your speech. This section is titled "How the ideas went in". It covers three lenses: data framing and retention, credibility and emotion in the handoff, and lexical bridging between your own thoughts and the provided material.

The read identifies up to four decisive moments, each labeled Smooth, Rough edge or Clunky. For rough-edge and clunky moments it provides a "Perfect pivot", a spoken line in your own register you could have used at that point. A momentum note at the end names one habit to keep and one to change.

If the integration read is unavailable, the assessment says so in a single line rather than showing a placeholder. No local rule can judge transitions, so nothing is faked.

Coaching notes and history

Three coaching cards appear below the integration read, under the heading "Coaching notes". Each names a strength or a skill to develop, written from your measured numbers. A label on the section says whether the cards were phrased by AI or by a local rule: "AI coaching" or "local notes, no AI". Both use the same underlying data.

The four categories are Pace, Coverage, Timing and Delivery. A card marked "Strength" has a green border. A card marked "Develop" has an amber border. When the AI is unavailable, the rule-based cards say the same things more plainly.

Your session is saved to History as soon as the assessment loads. A chip in the top corner of the assessment shows the save state: "saving to History" then "saved to History". You can revisit your score, coaching notes and the ideas you wove in at any time from the History list in the sidebar.

Trial session limits

The trial includes one freestyle session. It does not reset. The session has a listening budget of 330 seconds and a 45-minute window from the moment you start. If the listening budget runs out, the arena tells you and the session ends its tracking. The assessment still scores whatever was captured before the budget expired.

Once your session is used, the freestyle page shows a notice and points you to Pro or Premium for unlimited sessions.

What can go wrong

If no transcript appears, the microphone is almost always the cause. The article Use freestyle drills covers the most common hiccups: a missing transcript, ideas that fail to load, and very short takes that score oddly. For microphone permission steps, see Allow or restore microphone access. For browser and engine compatibility, see Supported browsers, engines, and fallbacks. For billing or plan questions that affect your freestyle allowance, see Manage billing, invoices, cancellation, and refunds.