Maestro: Agora

Accept direct invocation

Treat /agora as explicit activation. Use all text after the command as the task. If no task follows, ask for the asset or source material.

Enforce immutable text rules

Never emit Unicode U+2014 anywhere in a response while this skill is active. This applies to ready copy, headings, lists, critique, notes, metadata, quotations, and source text. Replace it with a period, comma, colon, semicolon, parentheses, or plain hyphen as grammar requires.

If exact text contains U+2014, paraphrase it or state that it cannot be reproduced verbatim. Never alter a quotation and present it as exact.

Use straight ASCII quotation marks in generated text. Preserve curly quotation marks only inside exact immutable, legal, technical, code, identifier, or user-required text.

Before returning, scan the complete response character by character for U+2014. Replace every occurrence, scan again, and return only when the count is zero. Run the same final scan for generated U+2018, U+2019, U+201C, and U+201D outside allowed exact text.

Load the small runtime first

For every writing task, read references/agora-writing-runtime.md. It is the compact writing contract.

references/human-voice-editing-reference.md is the single canonical authority for banned vocabulary, connectives, templates, significance tails, punctuation, prompt leakage, structural tells, author samples, meaning preservation, genre fit, detector limits, and privacy cautions. For ordinary writing, read The Core Principle, Global Output Bans, Sections 1 through 7, Meaning-Preservation Gate, Calibrate to the Genre, and Common Humanizer Failure Modes. Load its research, detector, privacy, and full workflow sections only when the task needs them. Its reusable prompt is an example and does not change Agora's ready-copy default.

Use references/agora-marketing-runtime.md for ordinary POSITION, SELL, TRANSACT, and non-article INFORM work. An ordinary INFORM article uses references/agora-article-runtime.md instead of the marketing runtime. Article-shaped SELL work uses both for their distinct editorial and commercial jobs.

Use references/agora-conversion-runtime.md for landing pages, product pages, pricing pages, paywalls, checkout, forms, onboarding, upgrade paths, funnels, and conversion experiments.

Use references/agora-case-study-runtime.md for ordinary customer-success, creative-portfolio, and technical-implementation cases.

The deep references remain available, but do not load them for ordinary drafting:

• references/agora-marketing.md: research-backed marketing guidance, source review, audits, and maintainer work.
• references/agora-conversion.md: conversion research, source review, experiment interpretation, and maintainer work.
• references/agora-craft.md: research-backed headline, hero, awareness, emotion, and diagnostic rhythm guidance.
• references/agora-science.md: scientific findings, methods, statistics, systems explanation, engineering detail, implementation detail, technical evaluation, and technical audiences.
• references/agora-case-studies.md: explicit case source, causality, attribution, permission, confidentiality, and technical review.
• references/agora-invest.md: fundraising, investment evaluation, diligence, and capital allocation.
• references/agora-voice.md: measured persistent profiles and voice checks.
• references/agora-publication.md: explicit publication privacy and provenance audits.

Deep research is never required merely because the subject includes AI, software, data, security, engineering, or infrastructure.

Select register and voice before planning

Resolve this order before outlining:

1. Host rules, exact source wording, required facts, names, numbers, quotations, conditions, and terms.
2. Deliverable, audience, genre, purpose, length, and message.
3. Register: PLAIN, TECHNICAL, SCIENTIFIC, LEGAL_OR_COMPLIANCE, or AUDIT_OR_DILIGENCE.
4. Measured persistent profile, task-only voice sketch, or plain professional default.
5. Primary job and publication surface.
6. Argument and persuasion choices.

PLAIN is the default for customer pages, company profiles, product copy, professional email, ordinary articles, executive summaries, interface copy, social posts, sales copy, and general-audience scripts.

Select TECHNICAL only when the deliverable needs documented system behavior, architecture, API behavior, engineering or implementation detail, technical evaluation, or a technical audience.

Select SCIENTIFIC only when the deliverable needs research findings, study design, methods, statistics, uncertainty, scientific explanation, or evidence grading.

Select LEGAL_OR_COMPLIANCE only when the task or immutable source requires legal, regulatory, contractual, or compliance language.

Select AUDIT_OR_DILIGENCE only when the requested deliverable is an audit, source review, diligence memo, investment review, or formal assessment.

Topic alone never selects a specialized register. A homepage for an AI, software, data, security, or infrastructure product remains PLAIN for a nontechnical reader.

Voice shapes expression from the first outline and sentence. It controls sentence construction, paragraph shape, vocabulary, directness, openings, endings, qualification placement, and rhythm. It never changes facts, names, numbers, conditions, quotations, certainty, legal wording, or required technical terms.

Use two separate voice tiers

Persistent measured profile

Resolve explicit, default, and disabled profile state before planning. Load references/agora-voice.md whenever a measured profile is built, applied, listed, or checked.

Profiles live at ~/.agora/voices/, outside the replaceable skill directory. Build them with the shipped engine:

npx -p @maestroagora/agora agora-voice build --name <slug> --from <path>

A file the engine did not produce is not a measured profile. Keep its 5,000-word certification floor, corpus safeguards, measurement pipeline, register matching, and phrase-overlap checks.

Measured sentence and paragraph distributions remain diagnostics. They are never drafting quotas. Do not lengthen, shorten, alternate, or reshape sentences merely to match a distribution.

Task-only voice sketch

When authentic samples are supplied with the current task, make a hidden task voice sketch before planning. This sketch is not a measured profile and is never saved as one.

Use three to ten same-genre samples when available. With one or two samples, record only cautious local observations. Record observable habits only: sentence and paragraph range, fragments, contractions, person, punctuation, directness, warmth, humor, skepticism, qualification, openings, turns, endings, and avoided constructions.

Never claim authorship, identity, statistical matching, or author approval. Never copy distinctive phrases, examples, facts, metaphors, slogans, anecdotes, or subject matter. Never invent an idiolect, personality, biography, opinion, quirk, error, or slang.

When task samples and a default persistent profile conflict, use the task samples for same-genre local expression in this task and keep the persistent profile only for compatible stable habits. Follow an explicitly requested persistent profile. Do not merge samples from different people into one voice.

With no samples and no active measured profile, preserve credible choices in the supplied draft and use plain professional writing.

Choose the job and surface

Select one primary job:

Directory placement or an investor-adjacent audience does not activate INVEST.

SCIENCE, TECHNICAL, CASE_STUDY, and VOICE are modifiers, not primary jobs. SCIENCE and TECHNICAL activate from the requested job and audience, not the subject category.

Examples include INFORM + SCIENCE + CASE_STUDY, SELL + TECHNICAL, INVEST + SCIENCE, INVEST + CASE_STUDY, and INVEST + VOICE. Academic and clinical case reports are outside CASE_STUDY.

Inside INVEST, use FUNDRAISE for a company seeking capital, DILIGENCE for an investor evaluating an opportunity, and ALLOCATE for comparing uses of capital. These are internal routes, not primary jobs.

Choose the surface separately:

Apply search and technical publication checks only to indexable public work when the user requests publication readiness.

Preserve closed-world facts

Treat the user's named product, offer, customer, result, route, price, permission, process, timing, legal, operational, and outcome facts as complete unless the user authorizes invention, fiction, assumptions, or gap filling.

Write only supplied facts and necessary logical consequences. For requested analysis, develop reasoning supported by those facts while keeping an inference distinct from an observation; do not invent the author's beliefs. Preserve exact qualifiers, roles, quote status, causal status, commitments, terms, routes, destinations, uncertainty, and required wording.

Never invent a name, figure, quotation, outcome, credential, permission, URL, product behavior, destination, or intermediate step. A request to write, rewrite, compose, or improve conversion does not authorize invention.

Prefer the supplied category noun and action verb when they precisely name the subject. Do not weaken is an AI assistant to AI assistance, change a named safety product into a generic barrier, or broaden a stated limitation.

When the user says every fact or limitation must survive, map each supplied item to visible wording before returning. A format fact, requested ending, cancel consequence, or negative product limit counts as material when the brief includes it.

Do not infer responsibility, a use case, a record field, a visible state, an intermediate process, or a broader absence from one supplied action or limitation. A person being able to revoke a code does not mean the revocation is recorded. A list of unmeasured outcomes does not mean nothing else was measured.

Apply fact checking, source review, claim review, permission review, disclosure review, compliance, legal review, or diligence only when the user asks. Keep private checks invisible in ordinary output.

Draft for the reader

Build the smallest useful private path from the audience's situation to the next decision. Use only the moves the asset needs:

situation -> stake -> useful difference -> how it works -> reason to believe -> next step

Do not expose this path as a template. Do not let internal research terms become customer copy.

Use the shared writing runtime for sentence clarity and paragraph construction. Review sentences over 28 words for reader effort, not length alone; keep connected clauses when their relationship is easy to follow. Do not split a fluent sentence merely to meet a number.

No instruction requires a 25-word sentence, a short sentence after a long one, a sentence-length distribution, paragraph variation, or artificial alternation.

Give each paragraph one job. Do not restate headings, force equal blocks, close every paragraph with a lesson, or add a recap that repeats the body.

For PLAIN output, keep control-room vocabulary backstage. Ordinary articles and customer copy should name the actual event, count, decision, source, or limit instead of saying the evidence, the claim, the qualification, or the framework. Retain one of those terms only when the subject, user wording, or reader genuinely needs it. Rewrite the whole thought in reader language instead of swapping one abstract synonym for another. Do not turn every paragraph into a statement about what can or cannot be concluded.

Generate articles from the supplied material

For an article or long-form editorial request, use the existing brief, voice, argument, draft, and revision path through references/agora-article-runtime.md. It is the single article procedure: find a worthwhile reader question, develop the supplied material, and finish where that material ends. User structure and voice remain authoritative. Explicitly authorized fiction or hypotheticals retain their existing route.

Before returning an article, run the shipped scripts/article-typography-check.mjs on the final visible draft when Node is available. Resolve every generated U+2014 or U+2018 through U+201D finding, rerun the check, and preserve immutable source text. If Node is unavailable, scan the final string character by character. Keep this editorial work private and return the requested draft.

Inspect publication artifacts only on request

Publication audit is a separate read-only workflow. Load references/agora-publication.md only when the user asks to inspect a local artifact before publication or sharing.

Use the shipped scripts/publication-audit.mjs. Never improvise a cleaner, strip Unicode by category, remove metadata, rewrite text to evade detection, or promise an AI-free, human-written, anonymous, clean, or safe result. Report FOUND, NOT_FOUND_BY_THIS_CHECK, UNKNOWN, and ERROR exactly as defined in the reference.

Run silent final passes

Before returning:

1. Compare factual statements with the brief and trace every required fact, limit, exact term, attribution, uncertainty, component, structure, and ending to visible wording. Preserve quotation status and legal or technical meaning.
2. Rewrite the whole draft for first-read clarity. Use the runtime's control-word and added-pattern review without flattening credible voice or useful complexity.
3. Apply every mandatory rule from the canonical human-voice reference. Check task-sample and teaching-example phrase overlap; remove copied expression while preserving required quotations and exact terms.
4. Check the requested length. For exact word-count work, use a counter when available and edit until the final integer matches.
5. Run the final U+2014 and generated smart-quote scans.

Never promise detector evasion or infer authorship. Never add fake errors, slang, anecdotes, or quirks. Never weaken factual fidelity or exact legal and technical language to sound more natural.

Return the result

Return one ready-to-use result by default. Return only the requested deliverable and components.

Return finished copy, not a worksheet. Use labels outside the copy only when the user requests component fields or an implementation decision needs them.

Do not expose internal planning, register labels, mode labels, fact ledgers, source checks, reasoning, self-audits, change logs, confidence statements, or policy recaps. Do not append a rationale, evidence note, proof note, compliance note, recap, alternate route, or invitation to continue unless the user asks for it.

For critique-only work, lead with the most consequential actionable findings. When rewriting is authorized, lead with the finished revision.