prospeak

Sound like a native of the room. Thirteen professional registers, five intensities, two languages, and it won't smell like AI.

 

View Demo · Report Bug · Request Feature · PlayMind Studio

Previously corpspeak. The name changed because the skill outgrew it: on a public tender or a clinical report it now actively forbids corporate jargon. Same dials, same defaults, same behaviour. Nothing you already used broke.

---

📑 Table of contents

• What is prospeak
• See it in action
• The five levels
• The thirteen branches
• One text, several worlds
• Why it's different: the anti-AI filter
• Bilingual, and tuned for Italian
• Install
• Usage
• How it's built
• How it was tested
• Roadmap
• Contributing
• Author
• Acknowledgments
• License

---

🤔 What is prospeak

Every trade has a way of talking, and an expert can tell in one sentence whether you belong to it. prospeak rewrites your text into the register of the world it is going to. Not one generic "professional" voice: thirteen of them.

The same sloppy input becomes "troubleshooting the pain points" in a management email, "si propone l'apertura di non conformità" in an engineering report, and "si trasmette in allegato la documentazione richiesta" in a letter to a public body. Same skill, different world, and it knows that in the last one the business jargon would have marked you as an outsider.

The contract first, because everything else depends on it: prospeak changes how something is said, never what is said. It never adds a fact, a figure, a date or a commitment that was not in your text, and it never quietly drops one. The length can move a little, because a register change costs or saves words, but this is not a summarizer and not an expander, and it is not there to correct your content. You write your ideas. It puts them in the language of the room.

Two dials, and that is the whole interface:

• How much register to add: five levels, from a light polish to full buzzword bingo.
• Which world to sound like: thirteen branches, inferred from your text if you do not name one, and switched per section when one text crosses several (a CV, a proposal).

It is also the anti-humanizer. Where a humanizer strips jargon out to make text sound natural, prospeak puts it in, on purpose, and then removes the structural tells that make writing smell generated. Most attempts at "make this sound professional" end up sounding like ChatGPT in costume: em dashes everywhere, fake triads, "stands as a testament to". This one keeps the jargon you asked for and takes the costume off.

Built with the Claude Code skill-creator. The original three level version was benchmarked head to head against plain Claude and won 95% to 70% on producing genuine register that does not read as AI.

✨ See it in action

Input (a flabby work message):

we tried the new tool but it didn't really help the team work faster, so we should probably stop using it

Level	Output
🪶 1 Light	We tested the new tool, but it did not improve the team's speed. I suggest we stop using it.
🗂️ 2 Business	The pilot did not produce measurable benefits on delivery times. I propose we discontinue it and reallocate the licences.
💼 3 Corporate (default)	Following the pilot, the new tool did not move the needle on team throughput. The recommendation is to discontinue adoption and reallocate the bandwidth to higher-impact solutions.
🎯 4 Executive	Pilot closed: no measurable impact on delivery times. Recommendation: sunset by month end, licences reallocated. Decision needed this week.
🔥 5 Extreme	Post-pilot, the tool failed to move the needle on team velocity. Recommendation: sunset adoption and reallocate bandwidth toward higher-leverage, ROI-positive enablers.

Input:

il pezzo che ci ha mandato il fornitore non va bene, è un po' fuori misura e adesso siamo in ritardo di due settimane

Branch engineering, level 3:

Il componente ricevuto dal fornitore presenta uno scostamento dimensionale rispetto alla tolleranza di specifica. Si propone l'apertura di non conformità e la definizione del trattamento, valutando anche l'accettazione in deroga per contenere l'impatto. Il ritardo stimato sul cronoprogramma è di due settimane.

Look at the verbs: si propone, not è stata aperta. A non conformità is a record with a number and a signature. The input never said one was opened, so the rewrite proposes it instead of declaring it.

Input:

vi mandiamo i documenti che ci avete chiesto, scusate il ritardo, se serve altro ditecelo

Branch public-sector, level 3:

Oggetto: procedura per l'affidamento di XXX, CIG XXXXXXX. Trasmissione della documentazione richiesta.

Con riferimento alla procedura in oggetto e in riscontro alla Vostra richiesta del GG/MM/AAAA, si trasmette in allegato la documentazione richiesta.

Si resta a disposizione per ogni eventuale ulteriore chiarimento o integrazione. Distinti saluti.

Note what did not happen: no quick win, no KPI, no roadmap. In a tender, business English is what marks you as an outsider. And the apology was dropped on purpose: in an act that gets protocolled, "scusate il ritardo" is a written admission on the record.

🎚️ The five levels

Level	Vibe	When to use
🪶 1 Light	"You, on a good day"	Little more than a context-aware corrector: takes the AI stink out, fixes the words that are wrong for that world, forces nothing.
🗂️ 2 Business	"The average good work email"	Ordinary office standard. Functional jargon only.
💼 3 Corporate (default)	"A competent manager"	Most cases. Jargon present but dosed and credible.
🎯 4 Executive	"A C-suite or a partner"	Dense, short, decisional. Answer first, then the number, then the ask.
🔥 5 Extreme	"A consultant who oversells"	Irony, motivational slides, buzzword bingo. Deliberately over the top.

You set the dial in plain language: make this corporate, light touch, livello estremo, da CDA.

Note the floor and the ceiling. Level 1 substitutes nothing: if "parlare con il team" works, it stays, and a text that was already clean comes back almost unchanged. Level 5 is deliberately loud but never ridiculous, and even there no factual claim is inflated: only the tone is.

Coming from the three level version? Light is still 1, medium is still the default, extreme is still the top. The scale doubled its resolution and levels 2 and 4 are the new steps in between. That extra resolution exists for one reason above all: dosing how much English goes into an Italian sentence.

🧭 The thirteen branches

A branch is the professional world whose register the text has to enter. It decides the vocabulary, the actors, the metrics that buy credibility, and the mistakes that expose an outsider.

Branch	For
corporate	Internal business communication, management, LinkedIn. The default.
tech	Releases, incidents, backlog, roadmap, product and startup
finance	Budget, margins, cash, valuation, banks, investors
consulting	Client recommendations, decks, steerco, proposals
engineering	Plants, sites, commesse, tolerances, testing, standards, safety
supply-chain	Purchasing, planning, logistics: lead time, stock, suppliers, deliveries
services	Client relationships under contract: scope, SLA, tickets, renewals
public-sector	Tenders, capitolati, formal acts, correspondence with a public body
legal-compliance	Contracts, privacy, compliance, risk
marketing-sales	Positioning, campaigns, funnel, pipeline, sales material
hr-people	Hiring, feedback, performance, internal announcements about people
healthcare-pharma	Clinical, hospital, pharma, medical device, market access
academic-research	Research claims, methods, papers, proposals, funded projects

You do not have to pick one. Say nothing and prospeak infers the branch from the format of the text, from the reader it addresses and from the vocabulary, in that order, then tells you what it chose so you can correct it in three words.

Not every branch climbs the same way

This is the part that keeps the skill from producing nonsense in technical worlds.

In rhetorical branches (corporate, consulting, marketing, services, HR), turning the dial up means more persuasive machinery and denser business vocabulary. In technical branches (engineering, finance, legal, public sector, healthcare, research), turning it up means more precision: units, tolerances, conditions, references to the norm or the method. Not more slogans. To an expert reader, inflated language on a technical document reads as an attempt to hide a weakness, so buzzwords there lower your standing instead of raising it.

Some branches also cap the dial. Legal and healthcare stop at 4. A public tender never goes out at 5. HR stops at 3 on anything addressed to one person about their job, because inflated language on bad news reads as contempt. When a cap applies, prospeak writes at the cap, says which level it actually wrote, and offers you the one you asked for.

🧩 One text, several worlds

A CV crosses three sectors. A proposal has a commercial part, a technical part and a legal part. Forcing one register on all of it produces the two failures this skill exists to avoid: a medical job described in startup words (nobody who worked there recognises it) or a software job described in clinical words (absurd).

So prospeak segments the text and picks a branch per section, on one rule:

The level belongs to the document. The branch belongs to the section.

One person is writing, at one intensity, about several worlds. The vocabulary moves, the voice does not. And the seams (the summary, the headings, the transitions) belong to the reader's world: in a CV aimed at a tech role, the profile at the top is tech, while the medical experience stays written in medical, because that is what proves you were really there.

Input:

2015-2019 ho lavorato in una ditta di apparecchi medicali, facevo assistenza tecnica agli ospedali e installavo le macchine 2019-2022 sono passato in un'azienda di logistica dove gestivo gli ordini ai fornitori e i magazzini dal 2022 lavoro in una software house come project manager, seguo i rilasci e il team di sviluppo

Output, level 3, three branches, one voice:

Dal 2022 a oggi, Project Manager, [azienda], software house Presidio del ciclo di rilascio e coordinamento del team di sviluppo: pianificazione delle attività, gestione delle dipendenze fino alla messa in produzione, allineamento con i referenti di prodotto.

2019-2022, Approvvigionamenti e magazzino, [azienda], settore logistica Emissione e gestione degli ordini fornitore e presidio delle giacenze: verifica della copertura rispetto al fabbisogno, interfaccia quotidiana con fornitori e magazzino.

2015-2019, Assistenza tecnica su dispositivi medici, [azienda], settore medicale Installazione e messa in servizio di apparecchiature medicali presso strutture ospedaliere, con interventi di assistenza tecnica sul campo.

The vocabulary moves, the rhythm does not. And no metric appears that you did not supply: on a CV, an invented percentage is the one that gets checked at the interview.

🧱 Why it's different: the anti-AI filter

prospeak borrows the detection rules from the humanizer skill and inverts the goal. The professional vocabulary you asked for (leverage, stakeholder, KPI, EBITDA, p95, non conformità) stays. The structural tells get removed every time. Eighteen of them, including:

• ❌ Em dashes and en dashes (the single most reliable AI tell)
• ❌ Decorative emojis, curly quotes, forced rule of three
• ❌ Fake "-ing" depth ("optimizing processes, enabling synergies, ensuring value")
• ❌ Significance inflation ("stands as a testament to") and motivational closers
• ❌ Invented substance, in both its forms: a percentage, a deadline or an article of law that was never in your text, and an act that never happened ("aperta non conformità", "dopo i confronti di [data]", "ci vediamo giovedì"). A placeholder replaces a value you have to supply, never the existence of a fact.
• ❌ Costume jargon: quick win in a tender, value proposition in an abstract, roadmap in a clinical report
• ❌ Negative parallelism ("non si tratta di X, si tratta di Y") and the same fact restated twice for emphasis
• ❌ The LinkedIn rhythm: humblebrag opening, one-line paragraphs, the closing question that begs for comments
• ❌ The zero rewrite (from level 2 up): if the only difference from your text is word order and two synonyms, the dial did not move and the job is not done
• ❌ Information loss: a duration you gave turned into a placeholder, a clause dropped because it was awkward, a fact lost to "compression". If the register really makes a sentence damaging, it can go, but you get told in one line, with the alternative

Then it runs a second pass, which is the one that matters: the false friends of the branch. The mistakes that expose an outsider to somebody who does that job every day. Confusing lead time with transit time. Calling an incident a bug. Quoting a relative risk reduction alone. Saying "il motivo è la performance" in a dismissal.

The principle in one line: real professional writing is dense but operational (actions, owners, deadlines, numbers). AI writing is dense but empty.

Fun fact: this README is em-dash-free on purpose. The skill practices what it preaches.

🌍 Bilingual, and tuned for Italian

Most "corporate speak" tools only handle English. prospeak is genuinely bilingual: every branch ships its lexicon in both languages, side by side, and by default it answers in the language of your input.

And it takes seriously the thing nobody else does: how much English to put inside an Italian sentence. Italian offices really do say "ti giro la mail", "ci allineiamo in call", "quick win", "mettere a terra". But the same word that makes you sound like an insider in a sprint review makes you sound absurd in a letter to a Comune. The rule prospeak applies is not "is this word used in Italy", it is "would this reader, in this document, say it".

So it sorts English words into three tiers (naturalised, professional, office slang), doses them by level and by branch, and gets the grammar right, which is where amateurs get caught: i budget not i budgets, le performance not le performances, fare un check before checkare.

It works in the other direction too, which is the half everyone forgets: the calques and false friends that mark an Italian writing business English. "Let's make a point on this", "I remain at your disposal", "actually" used to mean attualmente, "eventually" used to mean eventualmente.

🚀 Install

Option A, drop-in folder (Claude Code):

git clone https://github.com/giorgiozamboni/prospeak.git
cp -r prospeak ~/.claude/skills/prospeak

Option B, packaged skill: Download prospeak.skill from Releases and install it via the /plugin interface in Claude Code.

No API keys, no dependencies.

💬 Usage

/prospeak make this sound corporate, medium level:
"we lost two clients this quarter and we're a bit worried about cash"

Or just ask in natural language, with or without the slash command:

• "rendi più aziendale questa mail, livello leggero: ..."
• "buzzword it up for the board deck: ..."
• "riscrivilo per il CFO, livello executive: ..."
• "questa risposta al bando va scritta come si deve: ..."
• "make this postmortem readable for management: ..."
• "sistemami il CV, mi candido come project manager: ..."

If you do not specify a level, it uses 3 (corporate). If you do not specify a branch, it infers one and tells you which. If you do not specify a language, it matches your input.

🏗️ How it's built

Progressive disclosure: the skill loads only what a given request needs.

SKILL.md                      the router: two dials, the reconciliation, the process
references/
  levels.md                   the five levels, one input rendered at all five
  branch-detection.md         signals, ambiguities, writer vs reader
  multi-branch.md             one text, several worlds: segmentation, seams, the CV
  anti-ai-gate.md             18 structural tells, with before and after
  itanglese.md                how much English inside Italian, and vice versa
  lexicon-it.md               core Italian and itanglese lexicon
  lexicon-en.md               core corporate English lexicon
branches/
  _TEMPLATE.md                the contract: copy it to add a branch
  corporate.md  tech.md  finance.md  consulting.md  engineering.md  supply-chain.md
  services.md  public-sector.md  legal-compliance.md  marketing-sales.md
  hr-people.md  healthcare-pharma.md  academic-research.md

Every branch file follows the same contract: who speaks to whom, detection signals, bilingual lexicon tables, actors and roles, the numbers that buy credibility, bridge phrases, false friends, level tuning with an acceptance test, and a worked example.

🔬 How it was tested

Not by rereading it. By running it, with fresh agents that had never seen the design, then handing each output to an adversarial reviewer wearing three hats: an expert of that trade, an AI-tell detector, and a substance checker comparing input and output word by word.

It found things rereading never would. The biggest: the anti-invention rule covered invented numbers but not invented facts, so outputs would dutifully write XXX on every figure and then assert a non conformità that was never opened, a prior conversation that never happened, a meeting nobody scheduled. The root cause was worse: the worked examples inside the skill were teaching that behaviour.

Everything in the anti-AI filter above marked as an invented act, a negative parallelism or a zero rewrite exists because a test caught it.

🗺️ Roadmap

• Sector branches (thirteen, from tech to public tenders)
• Five levels instead of three
• Itanglese calibration, in both directions
• Per-section branch switching for texts that cross several worlds (CVs, proposals)
• An --explain mode that annotates every swap it makes
• A --check mode: score an existing text instead of rewriting it
• More languages (ES, FR, DE)
• Community lexicon packs for local office dialects

Have an idea? Open an issue.

🤝 Contributing

Contributions make the open-source world go round. The most valuable thing you can add is a false friend from your own trade: the mistake that makes an expert stop trusting the writer. New branches are welcome too (copy branches/_TEMPLATE.md). See CONTRIBUTING.md, open an issue, or send a PR.

And if prospeak made you smile (or saved your next deck), a ⭐ on the repo goes a long way.

👤 Author

Built by Giorgio Zamboni at PlayMind Studio, where we bring AI into real workflows, real decisions, and the work that actually matters.

• 🌐 Website: playmindstudio.it
• 💼 LinkedIn: giorgio-zamboni-engineer
• 🐙 GitHub: @giorgiozamboni

🙏 Acknowledgments

• The humanizer skill, whose AI-detection rules prospeak inverts
• Anthropic Claude Code and its skill-creator
• Wikipedia: Signs of AI writing, the source of the anti-AI patterns

📄 License

MIT. Use it, fork it, ship it. See LICENSE.

---