Skip to main content

ad-report

Draft, review, render, or verify a standalone analytical report from approved templates (Markdown, HTML, DOCX, PDF, slides, visualizations), preserving facts and asking before any export beyond the agreed local target. Use when the user wants a report from findings and evidence, or an existing analysis restructured for an audience.

Datos de origen

Repositorio
alexandremendoncaalvaro/agentic-development
Última actividad en el origen
25 de septiembre de 2026 a las 02:30
Idioma detectado de SKILL.md
inglés
Estrellas
2
Forks
0

Opciones de instalación

De forma predeterminada está seleccionado el prompt que primero revisa el origen. Puedes cambiar a un comando directo o descargar una copia local.

Revisa los archivos de origen

Lee SKILL.md y los archivos complementarios que muestra SkillsMP antes de decidir si quieres instalarlo.

Explorador de archivos
8 archivos

Mostrando SKILL.md

SKILL.md
Instrucciones de origen · Vista previa de solo lectura
name
ad-report
description
Draft, review, render, or verify a standalone analytical report from approved templates (Markdown, HTML, DOCX, PDF, slides, visualizations), preserving facts and asking before any export beyond the agreed local target. Use when the user wants a report from findings and evidence, or an existing analysis restructured for an audience.
summary
Draft or review standalone analytical reports from approved templates, then compose the available renderer and its visual and accessibility verification workflow.
<background_information> Codex and Claude Code apply this editorial workflow inline. The bundled script resolves and validates templates. Format-specific rendering remains with the available artifact capability rather than this skill. </background_information> # /ad-report Create or review one standalone analytical report from findings, evidence, and an audience. Normal use may write to an agreed local target but never mutates the template store or exports elsewhere without approval. <prime-directive> Return a self-contained decision artifact. Preserve every claim, name, number, date, citation, URL, identifier, commitment, uncertainty, exclusion, and requested action. Own the editorial plan, then compose the requested renderer and its normal verification contract. Never implement a parallel renderer or claim verification that did not run. </prime-directive> ## Step 1: Frame the report Accept findings and evidence, a brief, or an existing report. Resolve question or objective, audience, purpose, language, requested format, agreed local target, requested action, and hard constraints. Ask only for an ungroundable value that would materially change the report. Resolve optional language preferences through `ad-voice`. The conversation language controls direct owner communication and the first approval preview. The publication language is the default for outward text. An explicit language in the current request overrides the corresponding preference for that request only. When no language preferences exist, preserve the existing contextual inference. Build a silent invariant ledger. Read [references/source-policy.md](references/source-policy.md) before accessing any evidence beyond pasted text, an artifact attached for this report, a supplied public URL, and the resolved template. ## Step 2: Resolve the durable template Run from the consumer repository: ```bash node <skill-base-dir>/scripts/template-store.mjs resolve \ --bundled <skill-base-dir>/references/templates ``` Read [references/template-contract.md](references/template-contract.md). Select the most specific approved template matching the report type, destination, audience, purpose, and use conditions. Facts, safety, evidence limits, and format hard constraints outrank every template. Record the selected layer and shadowing; disclose a conflict only when it materially affects the report. When no approved owner or project report template exists, use the official-method bundled baseline and state outside the report, when material, that no private owner or team report corpus was authorized. ## Step 3: Build the editorial plan When the conversation language and publication language differ, build the complete initial report in the conversation language. Do not adapt it to the publication language before passing it through `ad-voice` in Step 4. Own the information hierarchy: self-contained executive conclusion, question or objective, method and evidence provenance, material findings, conclusions, recommendations or next actions, limitations, and optional depth or appendices. Keep findings distinct from interpretation. State uncertainty and evidence gaps where the reader uses them. Use a table, diagram, chart, or visualization only when it clarifies comparison, sequence, magnitude, or relationships. Do not add media as decoration or make a claim depend on color alone. ## Step 4: Compose owned capabilities Read [references/composition.md](references/composition.md). Invoke another skill only through its public contract: - `ad-ground` for a bounded method or implementation gap; - `ad-research` for an open question needing a durable study; - `ad-prism` when evaluation methodology, interpretation, or fit-for-purpose assurance is not already settled in the supplied findings; - `ad-voice` for the mandatory personal-voice and naturalization pass after facts, evidence limits, and structure are settled and before rendering; - `ad-template-tune` only on an explicit curation request; - the available DOCX, PDF, presentation, visualization, site, HTML, Markdown, or other artifact capability for the requested format. Every final editorial report must pass through `ad-voice`, even when no confirmed profile or external `humanizer` is available. Its public contract owns both the confirmed profile and the bundled naturalization fallback. Render only after this pass; a later textual correction must repeat it before re-rendering. For differing languages, `ad-voice` owns the two-stage language approval flow, including the first conversation-language preview, contextual adaptation, and final publication-language approval. Resume the same composed flow after the first approval; do not start a second voice pass or render the report between its two gates. Pass the settled editorial plan and content to the renderer. Respect every callee authorization, limitation, output, and gate; the stricter gate wins. Require its normal render, visual inspection, and accessibility verification. Return here for end-to-end factual and template coverage verification. If no renderer exists, return the editorial report in the safest usable available format and disclose the missing requested capability. Never fabricate a rendered file, screenshot, visual check, or accessibility result. ## Step 5: Verify privately After `ad-voice` returns the publication-language adaptation, independently run a meaning-preservation check against the approved conversation-language draft. List every changed, omitted, or added claim in the conversation language, then return the exact final publication-language text for approval before any outward export or publication. Read [references/evaluation.md](references/evaluation.md). Check the invariant ledger. Assign every applicable template rule and selected exemplar strength one internal verdict: `satisfied`, `improved`, `not applicable`, or `higher-priority override`. Revise unmet material items and re-run the composed renderer's checks after any content change that affects layout. After any textual revision in this step, repeat `ad-voice`, then render and repeat the invariant, template, visual, and accessibility checks before return. Any later text change invalidates both the voice pass and the rendered verification. An unresolved material gap is disclosed outside the usable report. "Same or better" is allowed only as traceable coverage plus a named target-specific improvement, never as a score based on length, recency, or visual polish. Do not return an artifact whose text changed after the most recent `ad-voice` output. ## Step 6: Return or export Return one usable report or the requested verified local artifact. Hide source selection, composition trace, and internal ledgers unless requested or needed for a material limitation. Writing to the agreed local target is part of the framed request. When that target is outside the repository or already exists, confirm the exact path with the owner before writing, and never overwrite an existing file without explicit approval. Before publishing or exporting beyond that agreed local target, show a receipt with the final artifact, exact destination, selected template ID and layer, format, completed and missing verification, material limitations, and exact action. Wait for explicit approval. Any content, format, or destination change invalidates the receipt. Report the resulting location or exact failure without retrying blindly.
Ver en GitHub