Agent-Skills.md

Agent Skills: Content Recast

|

UncategorizedID: ekson73/multi-agent-os/content-recast

Repository

ekson73/multi-agent-os
ekson73License: MIT

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ekson73/multi-agent-os/tree/HEAD/skills/content-recast

Skill Files

Browse the full folder contents for content-recast.

Download Skill

Loading file tree…

skills/content-recast/SKILL.md

Skill Metadata

Name
content-recast
Description
|

Content Recast

Overview

Re-cast a piece of your own technical content for a different target profile — {audience × abstraction × intent × language} — and hand off rendering to tools that already exist. Single responsibility = the semantic transformation (distill → recast → verify); rendering is delegated (composition-over-inheritance), so this skill never rebuilds editors. The quality differentiator vs. commodity "repurposing" SaaS is a mandatory faithfulness guard + information-loss note: a readable-but-inaccurate adaptation is worse than none (Devaraj et al., Evaluating Factuality in Text Simplification, ACL 2022).

For a fixed target — narrating a finished session/PR to a human (or agent) in a story-arc "Opera" register — use the specialised sibling skills/opera-debrief, which fixes the audience + register and adds dosed-intensity dials + a no-terror tone gate. Content-recast owns the general case (any audience × any register).

When to use

  • Re-target technical work for CEO / PO·PM / peer-dev / non-specialist / community / a daily standup.
  • Shift abstraction (low-level ↔ executive), intent (inform/convince/demonstrate/teach), or language (pt-BR ↔ en-US).
  • Produce a shareable artifact (deck, one-pager, NotebookLM podcast prompt) from raw technical notes.

When NOT to use: re-engineering the artifact itself (code/architecture migration — angular→flutter, monolith→microservices) → use a refactor/modernization agent. This skill can explain such a migration to an audience; it does not perform it. Pure documentation authoring → a docs generator. One-off rewrite with no profile shift → just edit inline. Narrating a finished session in the Opera register → skills/opera-debrief.

§0 — BEING > Rules (foundational)

This skill serves the operator's intent. If any phase/gate obstructs delivering value NOW, skip it, log Skipped <phase> — BEING > Rules, and proceed. HUMAN_DOMAIN (secrets · production PII · irreversibles · cross-org publish · cost) → escalate, never auto-act. Never invent facts to make an adaptation more compelling — that violates the faithfulness guard, which is non-negotiable.

Parameters

| Param | Default | Meaning | |---|---|---| | <source> (positional) | — (required) | Content to recast: inline text, a file path, or a glob. Empty → print usage. | | --to-audience | peer-dev | Who consumes it: ceo · po-pm · peer-dev · non-specialist · community · team-daily · free-text role. | | --abstraction | inferred from audience | executive (what/why/impact) · conceptual (how, no internals) · detailed (full technical depth). | | --intent | inform | inform · convince · demonstrate · teach · share. | | --lang | pt-BR | pt-BR · en-US (per the audience-design language policy). | | --format | md | md · slides · onepager · nlm-source · nlm-prompt · voice (render delegation — see Composition map). | | --readability | off | Optional Flesch-Kincaid grade target (e.g. 8). Treated as a check, not a guarantee (LLM readability control is imperfect — Farajidizaji et al., LREC 2024). | | --dry-run | off | Produce the neutral-brief + chosen lens + plan only; no recast/render. |

Argument parsing: the token(s) BEFORE the first -- form the <source>; every --key value / --flag after it is parsed as a parameter (matches the advertised <source> [--flags] syntax). If the input has no positional segment (starts with --, or <source> is implied), take <source> from the referenced file / prior conversation context. This avoids swallowing flags into the source.

Pipeline (0 → 6)

  1. Intake — parse <source> + params. Empty source → usage, stop. Resolve --abstraction from --to-audience if unset (ceo→executive · po-pm→conceptual · peer-dev→detailed · non-specialist→conceptual · community→conceptual · team-daily→conceptual).
  2. Distill → neutral brief — extract the source's atomic claims + supporting evidence + key terms into a compact, audience-neutral brief. This is the faithfulness anchor (the universal "distill-first, then rewrite" pattern). NEVER recast directly from the raw source.
  3. Select lens — pick the strategy lens from {abstraction × intent} (see Strategy lenses).
  4. Recast — rewrite the brief in the target profile through the lens: adjust register, depth, jargon, framing, and structure. Translate if --lang differs.
  5. Faithfulness check (MANDATORY) — verify every claim in the recast traces to a claim in the neutral brief (step 1). Flag/repair: unsupported insertions, distorted magnitudes, over-generalizations (generalization bias — PMC 2026), omitted load-bearing caveats. Emit an Information-Loss Note: what was simplified away or omitted, so the reader knows the boundaries (InfoLossQA, arXiv 2401.16475). If --readability set, score + (two-step) adjust + re-check.
  6. Render handoff — per --format, delegate to the producer (Composition map). Pass the recast content + a short rendering brief (tone/structure/visual hints). Never re-implement a renderer.
  7. Output — deliver the artifact (or producer-handoff instructions) + the Information-Loss Note + a one-line "what changed for this audience" summary.

Strategy lenses (abstraction × intent → technique)

| Lens | Fires when | Technique | |---|---|---| | Minto Pyramid | executive + (convince/inform) | Answer-first / SCQA; lead with the bottom-line + impact, support below. CEO/PO. | | Feynman | conceptual/non-specialist + teach | Plain language, analogy, no jargon; explain as if to a smart non-expert. | | Progressive disclosure | mixed audience / demonstrate | Layered: headline → key points → optional depth; reader chooses how deep (Cognitive Load Theory). | | Detail-preserving | detailed + peer-dev | Keep technical depth + precision; tighten structure, add signposting; minimal simplification. | | Opera (narrative-warm) | finished session/PR → human recap | Delegated to skills/opera-debrief (story-arc + dosed humour + moral + CTA + no-terror gate). |

Lens is a default suggestion — the agent may blend/override with rationale (cognitive-adaptation-freedom).

Composition map (render delegation — DRY, zero new editors)

| --format | Producer (delegated) | |---|---| | md | inline (this skill) | | slides | a slides producer (e.g. Gamma MCP) — fallback a pptx skill | | onepager | a PDF producer (e.g. make-pdf) — fallback a pdf/docx skill | | nlm-source | inline — emit clean markdown sized as a NotebookLM source | | nlm-prompt | inline — emit a ready prompt instructing NotebookLM to generate audio/video/infographic/slide/report/table from the nlm-source | | voice | bin/speak.sh --play via skills/voice-director — on-demand TTS (Gemini→ElevenLabs→Kokoro). Opt-in only, never a default (the operator may be where sound is unwelcome) |

Faithfulness guard (the differentiator — non-negotiable)

  1. Every recast claim MUST trace to the neutral brief (step 1). 2. No invented facts, numbers, or quotes. 3. No magnitude/causality distortion. 4. Load-bearing caveats survive or appear in the Information-Loss Note. 5. The note explicitly lists what was dropped/compressed. Research basis: ACL 2022 factuality, InfoLossQA 2024, FactPICO 2024, generalization-bias PMC 2026.

Output format

RECAST → <audience> / <abstraction> / <intent> / <lang> · lens: <lens>
<the recast content OR producer-handoff for non-md formats>

ℹ️ Information-Loss Note: <what was simplified / omitted / compressed>
🎯 For this audience: <1-line on the key reframing>

Anti-patterns (do NOT)

  • ❌ Recast directly from raw source (skip distill) → faithfulness drift.
  • ❌ Invent/embellish facts to be more convincing (AI slop) → violates the guard.
  • ❌ Rebuild an editor/renderer → use the Composition map.
  • ❌ Re-engineer code/architecture (Scope-B) → wrong tool.
  • ❌ Promise an exact readability grade → it's a check, not a lever.
  • ❌ Omit the Information-Loss Note when simplifying.

§Quality Tests (self-dogfood — 6/6)

  1. Self-Application — designed via research→converge→design; this SKILL.md is itself a recast of the research into the "AI-agent audience". ✅
  2. Non-Contradiction — composes (not duplicates) slides/NotebookLM/docs producers + opera-debrief; consistent with pre-creation scope-discipline / anti-theater / the language policy. ✅
  3. Survival — applied to itself it advocates faithful re-targeting; survives. ✅
  4. Bounded-Responsibility — --dry-run · Scope-A only · 5 formats · readability=check · DUED sunset. ✅
  5. Explicit-Exception — §0 BEING>Rules + HUMAN_DOMAIN escalation + cognitive-adaptation-freedom on lens. ✅
  6. Utility-Sunset — §DUED below. ✅ Pre-creation scope-discipline 6Q: 6/6 (WHERE · DRY=gap-confirmed vs the skill landscape + external SaaS · WHY=frequent need · WHO=operator+amnesic agents · FITS=content-lifecycle family · MIN=Goldilocks). Anti-theater 8Q REALITY: 8/8.

§DUED Sunset (qualitative, not counter-based)

Deprecate when ANY: the host provides a native faithful audience-recast primitive (E1) · operator retraction (E4) · ≥3 false-positive recasts where the guard misfires (E5) · a content-lifecycle family tool absorbs it (E6). Dormant-by-design otherwise.

Related Multi-Agent OS Artifacts

  • skills/opera-debrief/SKILL.md — the narrative-warm register specialisation: fixes audience (finished session → human/agent) + the Opera register + dosed dials + no-terror gate. Content-recast owns the general case; opera-debrief owns the "recap as a story" case. Shared faithfulness-gate lineage.
  • skills/postflight/SKILL.md · skills/morning-briefing/SKILL.md — fact sources a recast/debrief consumes (DRY).
  • skills/rule-quality-tests/SKILL.md — the 6 self-validity tests applied above.

§Refs

  • Genesis: forged via the agentic-tool-forge discipline (community).
  • Gates: pre-creation scope-discipline (6Q) · anti-theater grounding (8Q) · skills/rule-quality-tests (6 self-validity) · an audience-design language policy.
  • Research: ACL 2022 2022.acl-long.506 (factuality) · InfoLossQA arxiv 2401.16475 · FactPICO arxiv 2402.11456 · readability-control LREC 2024 2024.lrec-main.815 · ReadCtrl arxiv 2406.09205 · generalization-bias PMC 12042776.
  • Frameworks: Minto Pyramid Principle · Feynman technique · Progressive Disclosure · Cognitive Load Theory · plain-language/Flesch-Kincaid.
  • Producers: a slides MCP · NotebookLM · a PDF producer · pptx/pdf/docx skills.
  • Cross-link slug: [[content-recast]].

License

MIT (matches the multi-agent-os repo LICENSE).

Changelog

| Version | Date | Change | |---|---|---| | 0.1.0 | 2026-05-31 | Bootstrap. Scope-A communication re-targeting; 4-dim profile; distill→brief→recast→faithfulness-check→render pipeline; 4 strategy lenses; composition map (md/slides/onepager/nlm-source/nlm-prompt); mandatory faithfulness guard + information-loss note. Research-grounded. Forged via agentic-tool-forge discipline; 6/6 self-validity + 8/8 anti-theater + 6/6 scope-discipline. | | 0.2.0 | 2026-06-18 | Promoted user-scope → multi-agent-os (community). Layer-Purity scrub (org-specific strings removed) + relicense Apache-2.0 → MIT. Added the Opera (narrative-warm) lens row delegating to the new skills/opera-debrief sibling; added Related-Artifacts family section + License section. No pipeline/guard change. |