# Documentary Montage Production > Provider-independent workflow for AI agents producing documentary-style montages, archive-driven timelines, interview-supported sequences, nonprofit or advocacy shorts, historical explainers, and hybrid generated/archive edits. Use for editorial thesis development, research/source logs, fact-checking, archive rights, synthetic reenactment disclosure, chronology, interview selects, lower thirds, g… - **Type:** Skill - **Install:** `agentstack add skill-calesthio-generative-media-skills-documentary-montage-production` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [calesthio](https://agentstack.voostack.com/s/calesthio) - **Installs:** 0 - **Category:** [Developer Tools](https://agentstack.voostack.com/c/developer-tools) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [calesthio](https://github.com/calesthio) - **Source:** https://github.com/calesthio/generative-media-skills/tree/main/skills/production/content-formats/documentary-montage-production - **Website:** https://github.com/calesthio/OpenMontage ## Install ```sh agentstack add skill-calesthio-generative-media-skills-documentary-montage-production ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Documentary Montage Production Use this skill when the deliverable asks the viewer to believe, understand, or feel the significance of real people, real events, real institutions, or real-world claims. The central rule is: do not let the audience become more certain than the evidence allows. This is not legal advice. Treat rights, defamation, privacy, platform-policy, election, medical, financial, public-safety, and vulnerable-subject issues as escalation triggers for the appropriate human editor, client approver, platform owner, or counsel. ## Non-negotiable documentary posture Separate five things in every decision: 1. **Documented fact**: directly supported by a logged source. 2. **Corroborated interpretation**: a reasonable conclusion from multiple logged sources. 3. **Participant testimony**: attributed lived experience or opinion, not universal proof. 4. **Illustration or reenactment**: visual support that is not evidence. 5. **Editorial thesis**: the argument the piece makes from the above. If a shot, sound cue, texture, caption, edit, or generated image makes category 3, 4, or 5 look like category 1, revise it. ## Start with an evidence-first brief Before writing a script or making prompts, create a working brief with: - **Thesis**: one sentence the piece can fairly support. - **Audience and release context**: platform, geography, client, advocacy/commercial/newsroom/nonprofit context, and expected scrutiny. - **Core claims**: the 3-7 factual claims the piece depends on. - **Human stakes**: whose life, community, work, property, reputation, or safety is affected. - **Evidence map**: source status for each claim: verified, needs corroboration, contested, or must cut. - **Rights/disclosure map**: what footage, photos, voices, logos, documents, maps, music, and generated media will require permission, attribution, label, or review. - **Sensitivity level**: routine, sensitive, high-risk, or legal/editorial review required. Useful thesis test: "Because [documented change/conflict], [specific people/place/system] now faces [specific consequence]." If the thesis only says "this is shocking" or "you won't believe," it is not yet documentary-grade. ## Maintain a source and research log Keep a source log even for short social documentaries. A reliable log lets another agent or human reconstruct why each line and shot exists. Minimum fields: - `source_id` - title or description - creator/publisher/archive - URL, catalog ID, collection, box/folder/item, or file path - date created, date published, and date accessed - source type: primary record, official data, peer-reviewed research, news report, interview, participant-provided material, archive still/video, generated illustration, etc. - rights status: owned, commissioned, licensed, Creative Commons license, public domain claim, government work claim, fair-use review needed, unknown, or prohibited - provenance notes: original capture, digitized copy, upload mirror, C2PA/Content Credentials present, metadata stripped, chain of custody unknown - relevant claims or scene IDs - confidence and unresolved questions Minimum claim log: | Field | What to record | |---|---| | `claim_id` | Stable ID used in script, cards, and QA | | exact wording | The narration, lower third, card, or caption text | | evidence | `source_id` values and exact page, timestamp, table, quote, or frame | | support level | single-source, corroborated, disputed, inference, opinion | | on-screen treatment | spoken, card, lower third, graphic, caption, or omitted | | fact-check status | unchecked, checked, needs editor, needs counsel, cut | Do not cite a search result, social repost, or AI summary as the source for a consequential claim. Use the underlying original source when available. ## Fact-checking discipline Use an evidence hierarchy: 1. Original records, official data, court filings, archival catalog records, direct footage, and complete interview recordings. 2. Subject-matter experts, peer-reviewed research, institutional reports with disclosed methods. 3. Reputable journalism with named sources and corrections practices. 4. Participant testimony, clearly attributed. 5. Social media, unsourced compilations, AI summaries, and viral claims only as leads to verify elsewhere. For each claim: - Confirm that the source actually says the scripted claim, not merely something adjacent. - Check dates, units, denominators, names, locations, and causal language. - Prefer "after," "during," or "alongside" unless evidence supports "because." - Preserve uncertainty: use "about," "at least," "reported," "according to," or "records show" when appropriate. - Keep allegations attributed and give fair opportunity for response when making serious claims about identifiable people or organizations. - Remove or reframe claims that are true individually but misleading in sequence. For historical timelines, distinguish: - **event date**: when the event happened - **source date**: when the image/document/interview was created - **publication date**: when a source was published - **edit date**: when the montage was produced Cards such as "Chicago, 1968" must mean the image/clip actually depicts Chicago in 1968. If not, write "Archive image, Chicago, 1960s" or "Illustrative archival image; not the meeting described." ## Build chronology and causality carefully Documentary montages often compress time. Compression is allowed; false causality is not. Use one of these structures: - **Timeline spine**: chronological progression with dates or era cards. - **Argument ladder**: each beat proves a step in the thesis. - **Witness braid**: interview testimony alternates with records and archive to avoid one voice carrying all proof. - **Then/now contrast**: historical source followed by present-day consequence. - **Case-to-system zoom**: one person or place opens into broader data and policy context. - **Mystery/reveal**: acceptable only if the withheld information is not necessary for informed consent or safety. For every transition, ask: "Does this imply A caused B?" If yes, ensure the claim log supports causality. If not, add a card or narration bridge: "The records do not show why the decision changed, but they show what changed next." ## Handle interviews and audio selects Work from complete transcripts and source audio/video, not isolated clips. For each select, log: - speaker name, role, and release/consent status - source file and timecode in/out - transcript text, including meaningful pauses or corrections - context before/after the select - why the select is used - whether it is factual testimony, lived experience, opinion, or emotional color - cleanup applied: noise reduction, pause removal, filler removal, translation, subtitle edit Do not create "Frankenbites" that combine fragments into a meaning the speaker did not convey. If a sentence is assembled from non-adjacent moments, verify the combined meaning against the full interview and avoid hiding the join when the change is material. Use paraphrase in narration rather than manufacturing a cleaner quote. For translated interviews: - keep original-language audio audible where possible; - identify translator/subtitler status when high-stakes; - flag idioms, contested terms, and emotionally loaded phrasing for human review; - avoid synthetic voiceover that could be mistaken for the original speaker unless explicitly disclosed and approved. ## Archive and reference rights "Found online" is not a rights status. Treat every archive or reference item as requiring a rights decision before final delivery. Common statuses: - **Owned or commissioned**: confirm contract covers this use, territory, term, platform, paid ads, and derivatives. - **Licensed stock/archive**: keep license, invoice, allowed uses, restrictions, attribution requirements, and expiration. - **Creative Commons**: verify exact license version and restrictions. BY requires credit; NC may block commercial or fundraising use; ND may block edits/adaptations; SA may impose share-alike obligations. - **Public domain**: record why it is public domain in the relevant jurisdiction. A scan, upload, or compilation can add separate rights even when the underlying work is public domain. - **U.S. federal government work**: may be public domain under U.S. copyright law, but check third-party material, logos, privacy, personality rights, and agency-specific restrictions. - **Fair use / fair dealing / quotation**: do not self-approve for final. Prepare a rationale and escalate. - **Unknown or orphan**: do not use in final unless counsel/editor explicitly approves. Prefer a rights log tied to the source log: `asset_id`, `source_id`, owner/licensor, license terms, required credit, allowed platforms, edit restrictions, expiration, model/property/person releases, music/publishing/master rights, reviewer, decision date. ## Provenance and authenticity When available, preserve original filenames, checksums, embedded metadata, camera metadata, archive catalog IDs, and C2PA/Content Credentials. C2PA provenance is useful but not a truth guarantee: it can show signed provenance information and tamper evidence when supported, but it does not prove that the underlying claim is true. When metadata is missing or stripped, say so in the log; do not infer authenticity from visual appearance. ## Synthetic reenactment and generated media disclosure Use generated imagery, generated video, synthesized voices, face swaps, recreated documents, and reenactments only when they clarify a story without pretending to be evidence. Allowed uses, with labels: - "AI-generated illustration" for abstract or generic b-roll. - "Illustrative reenactment" for staged or generated depictions of reported events. - "Recreated from court filing / transcript / interview" when visualizing a record; show the record source if possible. - "Synthetic voice used for narration" when a generated voice is used. Avoid or escalate: - private person likeness without permission; - minors' likenesses; - public figures shown endorsing, confessing, acting in a crisis, or doing something not documented; - generated disaster, war, crime, protest, medical, or police footage that viewers could mistake for real footage; - fake authoritative sources, fake news footage, fake official documents, fake logos, fake camera timecodes, or fake archive stamps; - synthetic trauma scenes where the informational value is not necessary and approved. Disclose in three places when material is realistic or consequential: 1. **On-screen** near the generated or reenacted material. 2. **Credits/description** with a short generated-media note. 3. **Platform upload controls** where required. Platform and regulatory rules change. Re-check distribution rules at production time. Verified on 2026-07-11: YouTube requires creators to disclose realistic AI-generated or meaningfully AI-altered content during upload; TikTok asks/requires labeling for realistic AIGC and prohibits harmful misleading impersonation categories including certain uses of public figures, private adults without permission, and minors; Meta has used AI labels based on self-disclosure and industry signals across Facebook, Instagram, and Threads. Treat these as volatile facts. FTC and similar consumer-protection risk increases when synthetic media impersonates a business, government, official, endorser, expert, or ordinary person in a way that could mislead viewers in commerce. Escalate synthetic endorsements, fundraising appeals, political persuasion, product/medical/financial claims, and impersonation. ## Generated b-roll prompt translation Generated b-roll should translate evidence into clearly illustrative visuals, not invent new facts. Before prompting, write: - factual anchors: what must be true in the image/video - forbidden details: real faces, logos, badges, documents, locations, uniforms, timestamps, news chyrons, or readable text that are not licensed/verified - disclosure label: how it will appear on screen - visual role: mood, context, metaphor, process diagram, map, document texture, or reenactment Prompt pattern: ```text Create [illustrative / non-photoreal / stylized / clearly reenacted] b-roll for a documentary segment about [topic]. Show only: [verified or generic elements]. Avoid: real person likenesses, official logos, readable fake documents, news-watermark styling, camera timecodes, or details that imply this is authentic footage. Tone: [observational, restrained, sober, hopeful]. Composition: [specific shot/framing/motion]. On-screen label to be added in edit: [AI-generated illustration / illustrative reenactment]. ``` Example generated b-roll prompt: ```text Create a restrained, clearly illustrative shot for a documentary about a city housing waitlist: an anonymous stack of generic paper forms on a public-office counter, shallow depth of field, no readable text, no logos, no real addresses, no identifiable people, natural window light, observational tone. This will be labeled "AI-generated illustration" in edit. ``` Do not ask a model to create "archival footage," "news footage," "CCTV," "bodycam," "court exhibit," or "realistic video of [actual event]" unless the output is explicitly framed and labeled as reenactment and approved. ## Lower thirds, date/location cards, and source cards Lower thirds and cards are editorial claims. Fact-check them like narration. Use: - person lower third: `Name`, `role at time of event` or `current role`, and optional relationship to story. - date/location card: exact date and location only when known; otherwise `circa`, `undated`, `location reported as`, or omit. - archive/source card: `Source: [archive/publisher/collection]`, `Courtesy of [rights holder]`, or `Public domain: [basis]` when useful. - generated-media card: `AI-generated illustration`, `Illustrative reenactment`, or `Recreated from [source]`. Avoid: - "exclusive," "never before seen," or "declassified" unless source log proves it; - official-looking seals, typewriter textures, or broadcast chyrons that imply institutional provenance; - role labels that prejudice the viewer, such as "fraudster" or "victim," before the evidence establishes that label. ## Archival texture ethics Texture affects truth. Grain, scratches, projector flicker, VHS noise, monochrome, aspect ratio, and handheld shake can imply age or authenticity. Use archival texture only when: - it comes from the actual source; - it is a clearly aesthetic treatment that does not alter meaning; or - the piece's style makes the treatment obvious and not evidentiary. Do not add fake archive wear, timestamps, watermarks, camera overlays, tape glitches, or "old film" processing to generated or modern footage in a way that implies it is historical evidence. Restoration is also editorial. Do not over-clean footage so that important context disappears; record major stabilization, colorization, speed changes, or frame interpolation. Colorized, upscaled, slowed, or AI-restored archive should be labeled when the change is material. ## Music and sound restraint Documentary music should support comprehension, not coerce belief. Guidelines: - Keep music below narration and testimony; preserve intelligibility first. - Avoid villain cues under unproven allegations or identifiable people. - Avoid swelling music exactly where a contested factual claim is made. - Use silence, room tone, and natural sound to let difficult material breathe. - Do not add screams, sirens, crowds, gunshots, crying, or disaster sounds unless sourced, licensed, and contextually accurate; synthetic sound effects that mimic real events require disclosure when consequential. - For advocacy pieces, separate emotional appeal from evidentiary claims. ## Captions, transcript, and accessibility Provide captions for prerecorded audio in synchronized media unless the video is clearly a media alternative for text. Capti … ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [calesthio](https://github.com/calesthio) - **Source:** [calesthio/generative-media-skills](https://github.com/calesthio/generative-media-skills) - **License:** MIT - **Homepage:** https://github.com/calesthio/OpenMontage Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-calesthio-generative-media-skills-documentary-montage-production - Seller: https://agentstack.voostack.com/s/calesthio - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.