Respeak

Respeak

Most "humanizer" tools strip a word-list of AI tells. Respeak is a drafting method instead: reproduce the attention a human writer pays before choosing words, re-speak the draft aloud, and verify that every fact survived. The name is the core move — if the spoken explanation is clearer than the draft, the explanation IS the draft.

Think, speak, and write for a real person in a real situation. Preserve truth, intent, terminology, and useful personality. Remove machinery the reader can feel.

Do not optimize for an AI-detector score. Optimize for the moment of understanding: the reader should know what happened, why it matters, and what to do or think next without having to decode the writer.

Use the human model

Natural prose is the result of natural attention, not decorative informality. A human communicator usually does five things before choosing words:

1. Locates the other person. What do they know, see, want, fear, or misunderstand right now?
2. Chooses a centre of gravity. What is the one thing this passage must make clear?
3. Follows the causal thread. Who did what, under which constraint, and what changed as a result?
4. Anticipates friction. What honest question or objection will arise at this exact point?
5. Stops when the job is done. Do not add a recap, flourish, caveat, or call to action merely because prose convention expects one.

Use this priority order when goals compete:

1. truth and factual integrity;
2. reader comprehension;
3. continuity with the surrounding experience;
4. the author's or product's voice;
5. elegance.

Never sacrifice a higher item to improve a lower one.

Establish the contract

Before rewriting, identify:

• the intended reader;
• the surface: message, article, lesson, UI, narration, marketing, report, or localisation;
• what the reader should understand, feel, or do next;
• the facts, claims, numbers, names, links, examples, placeholders, and constraints that must survive;
• any established terminology or house voice;
• three to five genuine voice samples when matching a particular author matters.

If the source is ambiguous, do not silently resolve it through prettier wording. Preserve the ambiguity or ask for a ruling.

Work in four passes

1. Build a proposition ledger

List every proposition that must remain true. Include actors, causality, chronology, quantities, scope, uncertainty, and the requested action.

Separate content from style:

• Content may not drift without permission.
• Style may change freely when meaning, emphasis, and register remain intact.

Never invent quotations, statistics, examples, sources, user reactions, lived experience, certainty, or emotion. Humanity comes from specificity about what is true.

2. Re-speak the draft

Explain the same point aloud to a smart person in the target audience. If the spoken explanation is clearer than the draft, use the explanation as the new draft.

Do not perform a synonym swap. Rebuild awkward passages from their meaning.

Use these tests where relevant:

• Novice-question test: What would a capable newcomer ask after each sentence? Answer necessary questions where they arise.
• Storyboard test: Can a concrete claim be pictured, demonstrated, or traced? Replace empty adjectives with the event, object, behaviour, or consequence itself.
• Read-aloud test: Would a real person in this role say this sentence in one breath?
• Cold-read test: Can someone who skipped earlier context still understand standalone UI, notices, errors, and calls to action?
• Context test: Does every reference match what the reader can currently see or know?

When reasoning through an explanation, prefer this natural sequence:

situation → relevant constraint or rule → action → immediate consequence → why it matters

Omit any beat the reader already knows. Add a beat when its absence would force the reader to guess.

Speak like a person

Writing for speech is not the same as making written prose casual.

• Address one listener, even when the audience is large.
• Put one new mental move in each breath when the idea is unfamiliar.
• Use contractions and direct syntax when they fit the speaker.
• Name the actor before the action when several actors are present.
• Keep cause close to consequence; a listener cannot scroll back mentally.
• Prefer words that sound natural aloud over compact label-language.
• Let punctuation mark breath and thought, not visual sophistication.
• In narration, describe the beat that is happening now. Do not race ahead to the result.
• Do not pretend to share a physical or emotional moment the speaker did not experience.

Written prose may carry denser reasoning. Spoken prose needs cleaner chronology, shorter dependency chains, and fewer parenthetical detours.

3. Remove mechanical signals

Fix only the signals that are actually present. Do not sand the prose into generic plainness.

Structure

• Cut throat-clearing introductions, recap endings, and ceremonial conclusions.
• Avoid predictable context → three points → conclusion scaffolds unless the material genuinely needs them.
• Remove bold labels before every bullet, title-case heading overload, and decorative emoji.
• Do not announce the section's plan when you can simply do the work.
• Carry an argument with natural connectives rather than repeated role-labelled slots or a table disguised as prose.

Mechanical rhetoric

• Ration rule-of-three lists, question-and-answer setups, fake ranges, and contrast templates such as "not X, but Y".
• Keep at most one conspicuous rhetorical device in a short passage unless the author's samples justify more.
• Break em-dash chains into sentences when the dash is merely holding clauses together.
• Give inventories their own sentence instead of nesting them inside parentheses.

Rhythm

• Vary sentence and paragraph length without manufacturing staccato drama.
• Use contractions when the register is spoken or conversational.
• Leave a little harmless slack where a human breath belongs, but do not repeat a proposition as padding.
• Read the passage aloud. If the next beat is predictable, change the rhythm rather than decorating the vocabulary.

Vocabulary

• Prefer is, has, and concrete verbs over inflated substitutes such as serves as, features, or boasts.
• Remove stock transitions and fashionable abstractions when they add no meaning.
• Repeat the correct technical noun instead of synonym-cycling.
• Hedge feelings, interpretations, and uncertain evidence. Do not hedge settled facts.
• Let jargon earn its shorthand: define it before relying on it.

Voice

• Remove vague authority, sycophancy, promotional adjectives, unsupported certainty, and unnecessary disclaimers.
• Do not add chatbot closers, meta-commentary, or offers that the surface does not need.
• Do not personify systems, rules, data, or objects unless the voice deliberately calls for it.
• Preserve distinctive images, idioms, humour, bluntness, and imperfections that genuinely belong to the author.
• Match the author's measured habits rather than a generic idea of "natural".
• Preserve useful texture. A small asymmetry, blunt phrase, unexpected image, or quiet joke can carry more voice than a page of polished variation.
• Do not manufacture quirks. Idiosyncrasy is evidence from the author, not seasoning added by the editor.

4. Verify the rewrite

Diff the rewrite against the proposition ledger:

• Was any fact lost or added?
• Did an actor, cause, time, quantity, qualification, or requested action change?
• Did emphasis move?
• Are canonical terms and placeholders exact?
• Does the prose still sound like one person from beginning to end?

Then perform one stranger read. Identify the precise remaining tell and fix only that tell.

Teaching and explanatory prose

Teach as a patient expert speaking to an intelligent newcomer.

• Name the object before giving its rule.
• Introduce each concept before using its shorthand.
• State the normal rule before the exception.
• Give the reason beside the rule when the reason helps understanding.
• Anticipate the learner's natural objection and answer it directly.
• Use one concept per sentence when the logic is new.
• Use we for shared observation or calculation and you for the learner's knowledge or decision.
• Repeat a newly taught term when repetition teaches; cut restatements that merely rerun the same proposition.
• Put human stakes near important moments: what someone tries, notices, fears, gains, or loses. Use restraint.
• Let examples and visuals confirm the explanation, not replace it.

Apply the noob-question test sentence by sentence: "What would an intelligent beginner ask about what I just said?" Answer the necessary question now, not three paragraphs later. Do not answer every imaginable question; answer the one created by the sentence.

When a rule has a useful reason, unpack it rather than appending the reason to an already dense sentence. Short declaratives usually teach better than a compressed chain of conditions.

Bad pattern:

Do X unless Y, in which case Z, because Q.

Better pattern:

Normally, do X. Y changes that rule. When Y happens, do Z. This matters because Q.

Use contrast only when it resolves a real confusion. "Not X, but Y" is valuable when the learner is likely to believe X; otherwise it is rhetorical scaffolding.

For a sequence of lessons or chapters, run a cohesion pass with a knowledge ledger:

• record what the reader has already learned;
• introduce no cold terms;
• cash every forward promise;
• do not reteach the previous section under a new heading;
• keep one register across the whole sequence;
• ensure each section hands the reader naturally to the next.

A taught term keeps its taught form until the reader owns it. State the plain fact first, then attach the term. Do not suddenly switch to insider shorthand or a clever synonym.

Visuals, demonstrations, and dynamic state

Words must agree with the actual screen, image, animation, data, or event state.

• Refer only to objects the reader can currently see.
• Do not say above, here, watch, or this when the referenced thing is absent.
• Name the actor and action when several actors are possible.
• Describe a result only after it has happened; distinguish the current leader or provisional state from the final outcome.
• Use only information available to the reader at that moment. Never explain from hidden state or hindsight unless the surface is explicitly retrospective.
• In step-by-step narration, use one visible beat per sentence: actor, action, immediate consequence, then the reason if useful.
• Do not mention off-screen history as though the reader watched it.
• Distinguish an event that is merely possible, currently unfolding, provisionally true, and finally resolved.
• Make the words independently intelligible. The visual should confirm the mental model and bring it to life, not carry missing logic.

When prose and a visual disagree, verify the underlying fact before changing either. A vivid but incorrect explanation is worse than a plain one.

For a demonstration note, ask: "Have I explained what just happened, or merely named it?" A complete note normally identifies the situation, the governing reason, the action, and the consequence. Include the counterfactual only when it answers the learner's obvious "why not?"

UI and product copy

• Make every standalone string pass the cold-read test.
• Name the thing the user sees or the action they can take.
• For forced outcomes and errors, state the cause and the next useful consequence.
• Keep buttons short and unambiguous; do not make headings repeat the button beneath them.
• Preserve interpolation variables, markup, accessibility labels, length constraints, and platform terminology exactly.
• Ground product claims in evidence: a behaviour, number, artifact, or visible result.
• Let the interface do part of the explaining, but never rely on a visual that is not actually present on that screen.
• Keep headings, descriptions, controls, and feedback complementary. If two nearby elements say the same thing, give one of them a different job or remove it.
• Write empty, loading, error, completed, resumed, locked, and legacy states as full experiences, not afterthoughts.

For generated or interpolated copy, review the rendered sentence rather than only the template. Test zero, one, many, positive, negative, missing, unusually long, and translated values as applicable. Grammar belongs to the complete rendered string.

Marketing and professional prose

• Lead with the outcome or useful fact.
• Prefer one supported claim over three adjectives.
• Explain the mechanism when it creates trust.
• Keep calls to action proportionate to the evidence.
• Do not manufacture urgency, consensus, scarcity, testimonials, or social proof.
• For reports, separate observation, inference, recommendation, and uncertainty.

Calibrate confidence precisely:

• state verified facts directly;
• mark interpretations as interpretations;
• explain material uncertainty rather than hiding it in vague hedges;
• never convert "possible" into "likely", or "likely" into "certain", for a stronger sentence.

Localisation and multilingual copy

Treat each language as a native sibling, not a word-for-word mirror.

• Preserve the same propositions, actors, quantities, and consequences.
• Use native syntax, idioms, register, and established domain terminology in each language.
• Never transplant an idiom merely because its literal translation is understandable.
• Gloss specialist terms once, then use them consistently.
• Run a side-by-side drift check after every edit.
• Allow register to differ where languages require it; do not allow facts to differ.

Output modes

Choose the smallest mode that serves the task:

• Silent rewrite: return only the revised text.
• Ruling round: show before → after pairs with one short reason for contested or voice-bearing lines.
• Audit: preserve the text and list precise problems with suggested fixes.
• Cohesion review: report continuity, terminology, knowledge-order, and voice problems across a sequence.

Do not append a change log or explanation when the user asked only for finished copy.

Work across a large body of copy

For a product, course, book, documentation set, or campaign, local polish is not enough. Build a small authority stack:

1. Fact source: the implemented behaviour, primary evidence, or approved specification.
2. Voice source: accepted examples and explicit register decisions.
3. Knowledge or context ledger: what the reader knows and sees at each point.
4. Decision ledger: contested lines, rulings, reasons, and exceptions.
5. Current full-copy snapshot: the exact text readers encounter after all changes.

Read the complete experience in its real order. Search-based audits catch terms and patterns; they do not prove flow. Screenshots prove a frame; automated tests prove bounded facts; neither proves that a human can follow the experience.

Freeze passages that have passed review. Reopen them only for a factual conflict, a cross-surface inconsistency, or an explicit new ruling. Endless global polishing destroys voice and creates regressions.

Bank decisions

When the user accepts or rejects a meaningful line, record:

• original wording;
• accepted wording;
• why it won;
• surface and register;
• any exception.

Positive examples teach voice better than prohibition lists. Store project-specific terminology, tone, and rulings in a separate overlay; keep this core skill domain-agnostic.

When a user rejects a line, extract the general reason without turning one preference into a universal ban. Record whether the decision concerns truth, clarity, continuity, register, rhythm, or personal taste.

Verification ladder

Use the cheapest check that can genuinely prove the claim, then escalate when needed:

1. Static audit: banned terms, placeholders, formatting, duplicates, and terminology.
2. Proposition diff: facts, actors, causality, chronology, scope, and uncertainty.
3. Rendered-state check: full templates, layouts, line wrapping, accessibility, and dynamic values.
4. Sequential read: continuity, pacing, knowledge order, and voice across the whole journey.
5. Human-use check: can a representative reader understand and act without the author explaining beside them?

Do not claim a higher gate from a lower form of evidence.

Make it yours

Respeak stays domain-agnostic on purpose. Put your own voice rules — house terminology, banked phrasing decisions, per-project registers — in a separate overlay skill or ruling log that stacks on top of this one and wins conflicts. Positive examples of accepted lines teach an editor your voice faster than any list of bans.

Final gate

Do not ship until all are true:

• every source proposition survives;
• no new claim or lived experience was invented;
• terminology and placeholders are intact;
• the register fits the author, surface, and reader;
• references match visible and available context;
• the text reads naturally aloud;
• the whole piece feels written by one person;
• the reader never needs hidden context to understand a standalone surface;
• confidence matches the evidence;
• no remaining sentence is merely performing "human".