What is PunkVoice, the Spark Program application that analyzes LinkedIn drafts without writing them?

PunkVoice is the professional writing application yeshcube develops within Spark Program, and the first commercially available piece of that line. It operates on LinkedIn: its assistant, Sheena, works with an account’s profile, declared pillars and publishing history, checks every draft against that material and returns objections, questions and signs of inconsistency. The writing of the text and the decision to publish stay with the person.

The ERL scale (Evidence Readiness Level) measures the evidence maturity of a solution across six levels, from the initial hypothesis (ERL-0) to consolidated impact (ERL-5).

The problem: fluency without judgment

A generative model drafts a correct professional post in seconds, with an orderly structure, a predictable rhythm and a competent closing line. The marginal cost of producing text tends to zero, and with it goes the selection that cost used to impose: writing left out whatever did not merit the effort of being written.

When the tool takes on the writing entirely, the person hands the system the operations that hold judgment together: choosing the subject, settling on a position, ordering the argument and deciding what is defended in public. Spark Program describes that displacement as cognitive offloading and takes it as its object of study, on the observation that the effect of generative AI on learning depends on how the interaction is designed and on how much participation the person keeps.

On LinkedIn that displacement has a consequence visible in the feed itself: posts by different authors sharing structure, cadence and opening formulas. The professional reputation built there depends on someone’s judgment being recognizable, and a text that anyone could have signed does not serve that purpose.

Design requirements

Five requirements follow from that problem, and they govern PunkVoice’s design:

• The person’s authorship: the system intervenes in the writing process and the final text stays the responsibility of whoever signs it.
• Productive friction: the application returns the objection when it forces a rethink or calls for a decision to be justified.
• Contrast with the history: every draft is read against the positions the account has already published.
• Bounded scope: three to five pillars set the account’s subjects and filter out what falls outside them.
• Publishing outside the system: the person publishes the text on LinkedIn, with the application never acting on their behalf.

Architecture: the account state and the analysis layer

What separates the application from a conversation with a general-purpose model is the state it keeps. The account accumulates the declared profile and positioning, the pillars, the posts and comments retrieved from LinkedIn, the draft library with its analysis, the metrics and the successive assessments. That material is the condition for the contrast: with no history there is no inconsistency to detect, and an isolated session can only judge the text in front of it.

Sheena operates as an analysis layer over that state. It detects clichés and borrowed formulas, puts questions to an idea, connects a draft with earlier positions and flags inconsistencies before the text goes out. Generation relies on third-party generative models, which the application reaches through a gateway and with data minimisation: it sends the draft or the query and the profile context the feature needs.

Every action involving artificial intelligence draws on the PunkVoice credit system and shows its cost before running. That visible accounting serves a design purpose: turning to the model is an explicit decision at each step, with its price in view, rather than something the editor does by itself.

How an account works

The path has four steps. The first defines the base: the person gives Sheena their background, their positions and the audience they address. The second sets the pillars, three to five subjects that bound everything produced afterward and order the editorial calendar, where each idea is planned by concept, format and date, with states of planned, in progress, published or dropped.

The third is the writing. The workshop is a live editor that checks the voice rules as the text is drafted, counts the words and examines how it opens. It offers three feedback modes, Mirror, Challenge and Workshop, ranging from a reflection of the text itself to a direct objection to the argument. Before and after writing, the content vitamins bring together readings, podcasts and references on the account’s pillars.

The fourth step is cumulative. Every published text enters the history and becomes part of the material the next one will be read against, so a contradiction with an earlier position surfaces before publishing rather than after.

Internal measurement: Master Score, milestones and levels

The application summarizes an account’s evolution in a Master Score with five dimensions: consistency, autonomy, diversity, depth and impact. Levels correspond to bands of that score, and milestones are concrete goals the person either accepts from a rotating list or defines themselves.

Those indicators describe how the account behaves inside the product: how regularly it publishes, how work is spread across pillars and formats, how much intervention each piece requires and how far each text travels. The autonomy dimension comes closest to the program’s objective, because it records how much support the person needs to close a draft. None of the five amounts on its own to a measure of writing ability outside the tool.

Data and privacy

The data regime starts from a written commitment: account content is not used to train artificial intelligence models, whether its own or third-party ones. The account keeps the profile, the posts, the metrics, the milestones and the assessments because they are the material of the analysis, and deletion is requested from the settings.

PunkVoice also states the known limits of the models it uses: they can assert incorrect facts confidently, reflect biases in their training data and lack guaranteed recent knowledge. Their output is indicative, and checking any published fact falls to the person. The details are in the PunkVoice AI transparency commitment.

Evidence and validation status

PunkVoice holds level ERL-3, initial transfer: the application is commercially available and in use outside the research setting, with its data regime and the limits of its models publicly declared. ERL-3 is the threshold the hub requires before transferring any solution.

Level ERL-4 (validated scaling) requires longitudinal follow-up, studies across multiple contexts with cultural adaptation, peer-reviewed publication and cost-effectiveness assessment. That longitudinal follow-up is precisely the test Spark Program has to pass: establishing whether the person sustains outside the application the strategies practiced with it. Producing a better text while using it does not by itself demonstrate that transfer. It will be declared once the evidence is available.

What it is for and what falls outside

PunkVoice is intended for accounts that publish steadily with a professional objective and need to hold a recognizable position over time. The support covers defining editorial judgment, planning the calendar and reviewing each draft against what has already been published.

Automated writing falls outside the approach: the application does not produce the post in the person’s place and does not publish it on LinkedIn on their behalf. Reach as an end in itself falls outside it too, because the product works on authorship of the content and the coherence of the account.

PunkVoice carries Spark’s three principles into professional writing: activation rather than substitution, productive friction, and progression toward autonomy. The program runs within Allies and can take in external applications that share that approach. The details of the product are on the PunkVoice site and those of the line behind it in the Spark Program piece.