@maestroagora/agora 1.0.1 → 1.1.1

package/skills/agora/SKILL.md

@@ -1,59 +1,158 @@
----
-name: agora
-description: Write, rewrite, shorten, critique, or plan evidence-led marketing and sales content using the Agora argument-first method. Use when invoked as `/agora` and for marketing and sales copy; CTAs and microcopy; landing, product, and comparison pages; email and direct outreach; mobile-app onboarding, upgrade, and paywall screens; ads and social posts; editorial content; and spoken audio/video scripts plus written derivatives such as titles, descriptions, transcripts, captions, show notes, and companion pages.
----
-
-# 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 copy.
-
-## Load the authority
-
-Read [references/agora-marketing.md](references/agora-marketing.md) before producing or reviewing copy. Treat it as the canonical marketing-method reference. Preserve its evidence grades, limits, ethical controls, AI-writing-tell gate, and GEO/AEO boundaries.
-
-Treat current user-supplied facts, product documentation, live behavior, legal constraints, brand guidance, and approved claim sources as product truth. Never treat an example or marketing principle in the reference as authorization for a product claim.
-
-Apply the human-voice and AI-writing-tell gate to every Agora deliverable, regardless of brand, channel, privacy, or indexability. Treat the reference's banned vocabulary, connectives, phrase templates, significance tails, structural patterns, and typography as prohibited by default. Retain an item only when immutable verbatim material, verified proper names, required technical or legal language, or an explicit user or house requirement makes it necessary. Do not generate em dashes or curly or smart quotes in final copy; use straight quotes. Treat these as global `HOUSE` rules, not detector science.
-
-Apply CiteSurge-specific controls only when the request, named product, or active repository is CiteSurge. Ignore those controls for every other brand or app; use that project's verified rules instead. When giving engine-specific technical advice, recheck current official documentation because crawler and product behavior can change.
-
-## Route the asset
-
-Classify silently before working:
-
-| Class | Apply |
-|---|---|
-| `INDEXABLE_PUBLIC` | Agora argument, proof and public-claim review, human-voice gate, full written GEO/AEO gate, and technical publication handoff. |
-| `PUBLIC_NON_INDEXABLE_WRITTEN` | Agora argument, proof and public-claim review, human-voice gate, and written semantic/evidence rules. Skip crawl, canonical, schema, sitemap, and indexability checks. |
-| `WRITTEN_PRIVATE` | Agora argument, proof fidelity, human-voice gate, concrete entities, and self-contained claims. Skip public technical checks. |
-| `SPOKEN_ONLY` | Agora argument, proof gate, human cadence, breath, timing, and listener comprehension. Skip GEO/AEO formatting. |
-| `HYBRID` | Classify each surface separately. Treat spoken delivery as spoken-only; treat every published title, description, transcript, caption, show note, and companion page as written. |
-
-## Follow the workflow
-
-1. Capture the channel, primary audience, decision stage, problem or desire, current belief, one defensible destination belief, offer, one next action, verified proof, claim limits, voice, and format constraints. Infer only low-risk context. Ask only when missing information would materially change the audience, offer, claim, or action.
-2. Prioritize truth and safety, verified facts and brand constraints, buyer decision, clarity and citability, persuasion, brevity, then polish.
-3. Build the minimum complete argument. Select only needed moves: recognizable reality, real stakes, broken assumption, new criterion, reason to believe, difference and proof, tangible after-state, offer, and action. Keep one primary audience, destination belief, and CTA intent.
-4. Create emotion from a true situation, concrete consequence, and available agency. Preserve trade-offs and material conditions. Apply narrative, curiosity, fear, loss framing, guilt, scarcity, and social proof only within the reference's conditional controls.
-5. Attach proof to each material premise. Distinguish fact, inference, opinion, and promise. Keep source, date, scope, and limitation with the claim. Narrow or remove any claim whose proof is missing.
-6. Draft for the channel. Use the shortest complete decision path suited to device, intent, offer complexity, and risk. Keep one main job per block and one main claim per sentence. Do not force a long sales structure onto simple high-intent or transactional copy.
-7. Make the CTA the logical next step. Name one real action, predict the next screen or commitment, match intent, and add no new claim. Verify the destination or action exists.
-8. Apply written GEO/AEO and citability rules to every written asset: answer early when appropriate, name entities and scope, create self-contained evidence blocks, keep proof beside claims, add genuine information value, structure for reader intent, expose provenance, link contextually, and keep visible and machine-readable facts consistent. Add technical eligibility checks only for indexable public work. Never promise retrieval, selection, quotation, citation, ranking, recommendation, referral, conversion, or revenue.
-9. Skip GEO/AEO formatting for spoken-only delivery. Optimize spoken copy for natural cadence, breath, timing, emphasis, and comprehension. Apply written rules separately to any published derivative.
-10. Run the global competitor-swap, read-aloud, defense, continuity, and tell-density tests. Enforce the output bans: remove banned vocabulary and connectives without a narrow documented exception; prompt acknowledgements; stock phrase templates; significance tails; generated em dashes; curly or smart quotes; generated tripartite templates; `Not only... but also...` constructions; heading restatements; symmetrical blocks; templated bold-label bullets; decorative recaps; unneeded conclusions; fabricated human texture; and detector-evasion claims. Match supplied author samples without copying phrases or inventing biography, experience, motives, opinions, or mistakes.
-11. Compress in the reference's order. Preserve evidence, qualifications, legal terms, accessibility, and next-step clarity. Do not convert `HOUSE`, `C`, `EVIDENCE-INFORMED`, or `TEST` guidance into universal performance claims.
-12. Verify every final factual statement and CTA against supplied sources and current project truth. For CiteSurge work only, apply its claim-ledger, methodology, AI-writing scanner, GEO/AEO positioning, and release controls from the reference.
-
-## Enforce the claim gate
-
-Never invent or imply unsupported claims, features, prices, routes, evidence, deadlines, urgency, scarcity, testimonials, results, motives, pain, exclusivity, comparisons, or guarantees. Never conceal price, fit, risk, evidence, or material conditions to create curiosity.
-
-When required truth is unavailable, omit or narrow the claim. If omission would make the asset unusable, put a concise verification need after the ready-to-use copy; do not fill the gap with a guess.
-
-## Return the result
-
-Return ready-to-use copy first. Do not expose chain-of-thought, internal classification, planning scaffolds, or rule-by-rule reasoning. Add assumptions, source gaps, rationale, or tests only when requested or materially necessary.
-
-Default to one shortest complete output suited to the channel. Do not provide multiple near-duplicate variants unless requested. For critiques, lead with the revised copy when rewriting is authorized; otherwise lead with the most consequential actionable findings.
+---
+name: agora
+description: Write, rewrite, shorten, critique, or plan truthful persuasion and objective company positioning with the Agora argument-first method. Use when invoked as `/agora` or for marketing and sales copy; investor, funding, pitch, company-profile, category, or brand descriptions; CTAs and microcopy; landing, product, and comparison pages; email and direct outreach; mobile onboarding, upgrade, and paywall screens; ads and social posts; editorial or educational content; and spoken audio/video scripts plus written derivatives such as titles, descriptions, transcripts, captions, show notes, and companion pages.
+---
+
+# 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 copy.
+
+## Load the authority progressively
+
+Use [references/agora-marketing.md](references/agora-marketing.md) as the canonical authority. Load only the sections required for the current job:
+
+1. Always read `Core doctrine`, `The destination belief`, `Mechanism and company transformation`, `Commercial modes`, `Construction order`, `Proof and claim control`, and `Human voice and AI-writing-tell gate`.
+2. For `SELL`, `INVEST`, or `POSITION`, also read `Emotion without hype`, `CTA standard`, `Rejection gates`, and the closest factual pair under `Applied weak and strong pairs`.
+3. For written work, read `Written GEO/AEO and citability gate`. Load engine-specific boundaries only for engine-specific technical advice.
+4. For spoken work, read `Spoken delivery`. Apply the written gates separately to any published title, description, transcript, caption, show note, or companion page.
+5. For CiteSurge work only, read the CiteSurge-specific controls. Never apply them to another brand.
+6. Load the evidence register when making or reviewing a research-backed general claim. Do not load it merely to draft from supplied product facts.
+
+Treat current user-supplied facts, product documentation, live behavior, legal constraints, brand guidance, and approved claim sources as product truth. Examples teach structure, not facts.
+
+Resolve conflicts in this order: truth and safety; verified facts and brand constraints; the audience's decision; clarity and citability; persuasion; brevity; stylistic polish. Infer only low-risk context. Ask a question only when the answer would materially change the audience, offer, claim, or action; otherwise omit or narrow the unsupported point.
+
+## Choose the commercial job
+
+Select one primary mode before drafting. An explicit mode wins. Otherwise infer it from the audience, requested asset, and decision:
+
+| Mode | Infer when the task concerns |
+|---|---|
+| `INVEST` | Investors, funding, pitches, or capital-focused profiles |
+| `SELL` | Marketing, sales, ads, landing pages, outreach, or paywalls |
+| `POSITION` | Company profiles, category narratives, or brand descriptions |
+| `INFORM` | Editorial or educational work |
+| `TRANSACT` | Buttons, confirmations, alerts, or utility microcopy |
+
+Do not confuse mode with surface. An investor description on Crunchbase is `INVEST` on an indexable public, objective profile surface. A paywall is `SELL` on a written interface surface. A published transcript is written even when its source script is spoken.
+
+For `INVEST`, state investment relevance in the visible copy. Placement in an investor directory does not carry this move. Use a neutral sentence that says the company `merits evaluation because` of a specific supported mechanism or workflow position. Do not end with abstract category restatements built on `represents`, `focused on`, `approach to`, or `positioned at`. Never substitute traction, market hype, or a funding CTA.
+
+When several fields form one company profile, make the longest field carry the complete commercial spine. A short field may compress the problem or criterion with the transformation, but the Full Description cannot become a feature inventory. Do not include every supplied capability by default; omit facts that do not advance the destination belief.
+
+## Route the surface separately
+
+Classify each requested surface:
+
+| Surface | Apply |
+|---|---|
+| `INDEXABLE_PUBLIC` | Public claim review, written human-voice and GEO/AEO gates, plus technical publication handoff |
+| `PUBLIC_NON_INDEXABLE_WRITTEN` | Public claim review and written human-voice and semantic/evidence gates; skip crawl and index checks |
+| `WRITTEN_PRIVATE` | Proof fidelity, concrete entities, self-contained claims, and human-voice gate |
+| `SPOKEN_ONLY` | Proof gate, human cadence, breath, timing, and listener comprehension; skip GEO/AEO formatting |
+| `HYBRID` | Route every spoken and written derivative separately |
+
+Channel rules cap tone without deleting the argument. For objective platforms such as Crunchbase, write the strongest objective argument permitted and explain any channel conflict after the copy.
+
+For a tightly limited objective company profile, preserve the supported operational problem and company transformation. A category label is optional. Do not spend the limit on `Company is [category] that [feature list]` when the supplied facts support a problem-to-transformation argument.
+
+## Follow the exact construction order
+
+Use this order without moving compliance ahead of persuasion:
+
+`Commercial job -> destination belief -> argument -> proof gate -> channel formatting -> written GEO/AEO when applicable -> AI-writing-tell edit -> compression.`
+
+1. **Commercial job.** Choose one primary mode, audience, offer, decision, and real next action or investment relevance.
+2. **Destination belief.** State the narrow conclusion the audience must accept before the requested decision makes sense.
+3. **Argument.** For `SELL`, `INVEST`, and `POSITION`, build the mandatory commercial spine below. For `INFORM` and `TRANSACT`, use only the argument moves the task needs.
+4. **Proof gate.** Verify or narrow every premise before channel, GEO/AEO, style, or length work. Compliance cannot rescue a missing or unsupported argument.
+5. **Channel formatting.** Fit the objective, editorial, interface, outreach, spoken, or other channel without deleting load-bearing moves. Honor explicit length and format constraints when they do not require factual distortion.
+6. **Written GEO/AEO.** Improve extractability and verification only after the argument passes the proof gate.
+7. **AI-writing-tell edit.** Apply the global voice and output bans without changing facts, scope, causality, or uncertainty.
+8. **Compression.** Remove repetition and decoration last. Preserve the stake, mechanism, proof, qualification, destination belief, action relevance, and any explicit usable length range.
+
+## Build the mandatory commercial spine
+
+For `SELL`, `INVEST`, and `POSITION`, construct this exact sequence:
+
+`Recognizable reality -> consequence or opportunity -> new decision criterion -> company or product mechanism -> defensible destination belief -> action or investment relevance.`
+
+Moves may share sentences; none may disappear.
+
+Before delivery, silently locate the exact words that carry each of the six moves. If any move has no concrete phrase in the draft, rebuild it. A compact hero, profile, paywall, email, or spoken pitch is not exempt; combine moves instead of deleting them. Do not expose this audit.
+
+- **Recognizable reality:** Name the audience's current situation without inventing pain, motive, or urgency.
+- **Consequence or opportunity:** Make one truthful stake felt through a concrete business effect, decision risk, blocked task, or available gain. Count this move only when the draft names that effect, decision, task, exposure, or opportunity. Words such as `matters`, `confidence`, `better`, or `before acting` do not carry the stake alone.
+- **New decision criterion:** Establish what a credible choice must do, based on the supplied facts.
+- **Company or product mechanism:** Explain what the company changes, what causes the change, and what limits apply.
+- **Defensible destination belief:** Land the one supported buying, investment, or positioning conclusion.
+- **Action or investment relevance:** State the logical next step, or why the company merits evaluation when the channel does not permit a CTA.
+
+For an objective company profile, keep the spine factual and compact. Do not replace the stake with hype or replace action relevance with a sales command.
+
+## Enforce the proof gate
+
+Attach proof to the argument before applying channel rules or compliance gates.
+
+- Require every included fact to prove a premise, resolve an objection, or enable action. Remove facts that merely decorate the draft.
+- Keep fact, inference, opinion, and promise distinct. Preserve source, date, scope, qualification, and evidence limits where material.
+- Treat product behavior and user-supplied facts as evidence only for the claim they directly support.
+- Never invent or imply features, results, prices, routes, deadlines, traction, leadership, market size, savings, performance, urgency, scarcity, testimonials, comparisons, motives, or guarantees.
+- Do not turn an operational mismatch, exception, or review task into financial risk, a control deficiency, compliance exposure, loss, or business impact unless the supplied evidence supports that consequence. Keep the stake at the supported task or decision.
+- Treat `emerging`, `growing`, `established`, `crowded`, `large`, `expanding`, `increasingly`, `accelerating`, `rising`, `surging`, `rapidly`, and similar category-stage or market-trend language as claims. When no market evidence is supplied, name the category and current situation neutrally.
+- If proof is missing, narrow or remove the premise. Put a concise verification need after the ready copy only when omission would make the asset unusable.
+- Do not add a generic disclaimer such as `outcomes are not guaranteed` merely because a result claim is unavailable. Omit the result claim. Preserve a qualification only when it is material to the remaining copy or required by law, policy, or the user.
+- Do not turn crawler access, schema, content changes, or observed citations into promises of retrieval, selection, ranking, recommendation, conversion, or revenue.
+
+## Apply the rejection gates
+
+Reject and rebuild a `SELL`, `INVEST`, or `POSITION` draft when any condition holds:
+
+- It is only a category definition, claim-ledger paraphrase, or feature inventory.
+- An objective profile uses the shape `Company is [category] that [feature list]` even though the supplied mechanism supports a decision problem or company transformation. Start from that problem or transformation instead.
+- It has no felt stake, company transformation, or destination belief.
+- A commercial-spine move disappeared to satisfy a short or objective format.
+- A paywall merely repeats plan facts without naming the blocked action and how the supported upgrade change resolves it.
+- The facts do not prove a premise, resolve an objection, or enable action.
+- It passes a competitor-swap test because the named company has no factual mechanism or boundary.
+- It presents compliance, formatting, or citability as persuasion.
+
+Reject primary brand verbs such as `shows`, `helps`, `supports`, or `built for` when a supported causal verb is available. Use the strongest factual verb the evidence permits, such as `records`, `reconciles`, `routes`, or `turns`. Never strengthen the verb beyond the evidence.
+
+Run the proof, competitor-swap, continuity, CTA, read-aloud, defense, fidelity, and tell-density tests before delivery.
+
+## Preserve truth, ethics, and voice
+
+Create emotion from a true situation, a concrete consequence or opportunity, and available agency. Do not invent fear, loss, guilt, scarcity, social proof, exclusivity, or human texture. Preserve material trade-offs and conditions.
+
+Apply the human-voice and AI-writing-tell gate to every deliverable. Do not generate em dashes or curly or smart quotes in final copy. Remove banned vocabulary, connectives, phrase templates, significance tails, stock generated tripartite rhetoric, `Not only... but also...`, `not merely... but...`, and `not just... but...` constructions, decorative recaps, prompt acknowledgements, and fabricated informality unless immutable text or required legal or technical language makes preservation necessary. Use straight quotes.
+
+Before delivery, scan every generated character in the ready copy and any note after it. Replace generated U+2014, U+2018, U+2019, U+201C, and U+201D punctuation with an ASCII alternative, including when restating a user's unsupported phrase. Do not let a verification note reintroduce a banned mark.
+
+Run a separate triad sweep. Never generate the prose syntax `X, Y, and Z` or `X, Y, or Z`; treat this as a mechanical output constraint. Rewrite any sequence of three parallel nouns, adjectives, clauses, or sentence fragments. A flourish such as `measurement, traceability, and discipline` fails. A recap such as `evidence capture, prioritization, and human review` also fails; write `captures evidence and keeps prioritization under human review`. Recast even sourced facts with a natural pair or connected prose. Preserve the three-part syntax only inside immutable quoted or legal text, a proper name, or text the user requires verbatim.
+
+Scan the final prose from the last sentence backward. Find every generated `, and` and `, or`; when it closes a three-part parallel sequence, rewrite it. Check every occurrence, including the last sentence and any note. Do not stop after the first repair.
+
+Never promise detector evasion. Never change facts, names, numbers, dates, quotations, citations, causality, uncertainty, or legal meaning to sound human.
+
+## Preserve written GEO/AEO boundaries
+
+For written work, answer early when the format calls for it, name entities and scope, keep proof beside claims, build self-contained evidence blocks, expose provenance, add genuine information value, and keep visible and machine-readable facts consistent.
+
+Apply technical crawl, canonical, sitemap, metadata, structured-data, and delivery checks only to indexable public work. Treat engine rules as engine-specific and recheck current official documentation before technical advice. Never promise retrieval, selection, quotation, citation, ranking, recommendation, referral, conversion, or revenue.
+
+## Isolate CiteSurge controls
+
+Apply CiteSurge claim-ledger, methodology, scanner, GEO/AEO positioning, public-address, release, and review controls only when the request, named product, or active repository is CiteSurge. Ignore them for every other brand. For CiteSurge public work, verify current repository truth and required smokes before release.
+
+## Return the result
+
+Return ready-to-use copy first. Do not expose chain-of-thought, mode classification, planning scaffolds, private reasoning, or a rule-by-rule explanation. Add assumptions, verification needs, or an objective-channel conflict only after the copy and only when material.
+
+Run the literal punctuation sweep after composing any post-copy note, then return the result.
+
+Do not deliver until the trend-claim, contrast-template, and triad sweeps also pass the final copy and every post-copy note.
+
+Default to one shortest complete output suited to the channel. Do not provide near-duplicate variants unless requested. For critiques, lead with revised copy only when rewriting is authorized; otherwise lead with the most consequential actionable findings.

package/skills/agora/agents/openai.yaml

@@ -1,4 +1,4 @@
 interface:
   display_name: "Maestro: Agora"
-  short_description: "Evidence-led marketing across every channel"
-  default_prompt: "Use $agora to write concise, evidence-led copy for this asset."
+  short_description: "Truthful arguments for buyers and investors"
+  default_prompt: "Use $agora to lead one audience from a real stake to a defensible buying or investment belief, then return ready-to-use copy."