메뉴
Generate AI ad assets — Veo 3 video prompts and AI image static prompts — with brand context, compliance, and forensic direction. The output is prompt text the user pastes into the AI tool, not the final asset.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
How this brain saves, syncs, and clones itself on GitHub — the required procedure for EVERY git push, pull, fetch, clone, commit, or gh command in a Parker Brain repo. Covers the credentials (the setup_parker_brain tool, never the user's own login, never gh), submodules, merge conflicts, and the rare self-hosted exception. Use whenever saving work to GitHub, syncing or updating the repo, cloning the brain, fixing a failed push or auth error, or when asked to save, back up, or share the brain.
The front door to building a brand's brain. Run this when standing up a new brand — all it needs is the Parker MCP connected. It welcomes the person, provisions the brand's own repo through the setup_parker_brain tool (Parker MCP), clones that repo and works inside it, connects and tests the data, asks the short brand intake (the things Parker genuinely can't observe), runs the full build in dependency order, saves it to GitHub, and hands off to get-started. Follows prompts/onboarding-runner.md as the detailed method. Use when someone says set up a brand, build the brain, onboard a new brand, or get started building. Also the right entry when the brand's brain already exists — it detects a built brain and retrieves it, handing off to get-started instead of rebuilding.
The weekly check that keeps this brain current with the parker-brain standard — without overriding anything the team chose. Compares the pinned parker-system/ release against the newest factory tag and offers the bump as one short question; on a yes the migrations run and a script re-syncs the copied skills deterministically. What the team declines stays declined; what they modified stays theirs. Use weekly as a scheduled routine, or when asked to update the brain, check for updates, sync with the standard, or "what's new in Parker".
One-time setup for a fresh brand brain instance. Registers the standing routines (refresh-context, dream, harvest+evaluate ideas, research-loops, update-brain, self-improve) as scheduled cloud agents so they run on cadence without being asked. The job definitions already travel with this repo; this skill arms the schedules, which are per-account and can't be committed. Run once after cloning the brain into a new Claude Code cloud instance, or when asked to set up / register / schedule the routines.
Write headlines for ads — static, video text overlays, paid social, landing page hero. Picks the right structural pattern for the brand type (lifestyle vs problem/solution), the awareness stage, and the customer language available.
Write hooks for a specific ad — the first three seconds that earn the click and the watch. Uses the proven hook formats from the canonical taxonomy, picks the right one for the brand and audience, and produces hooks built around a specific ICP and emotion.
| name | ai-ad-generation |
| description | Generate AI ad assets — Veo 3 video prompts and AI image static prompts — with brand context, compliance, and forensic direction. The output is prompt text the user pastes into the AI tool, not the final asset. |
| triggers | ["veo prompt","veo 3 prompt","AI video prompt","AI video generation","text-to-video prompt","cinematic prompt for AI","multi-shot video sequence","storyboard to prompt","AI static ad","AI image ad","generate a static with AI","midjourney prompt for an ad","recreate this static","adapt this competitor static","AI ad creative","prompt for AI ad"] |
Produce the prompt text that goes into an AI generation tool — Veo 3 for video, image models for static — so the output matches the brand's voice, compliance, and visual language. Whatever the model fills in by default is the model's interpretation of the brand, which is almost never what the brand actually is. The skill's job is to remove that gap by being forensically specific about everything in the frame, on the soundtrack, in the copy, and in the visual identity.
This skill produces prompt text. It does not generate the final asset, evaluate the asset's performance, or iterate on a winning asset — those are different skills (ad-account-analysis, iterations).
The generation methods this skill runs on are canonical, not improvised. Before writing a prompt, load what parker-system/creative-strategy-context/expertise-routing.md names for creative generation: veo3-video-prompting.md for video, ai-static-ad-generation.md for statics, static-ad-recreation.md for a recreation, and visual-vocabulary-method.md so the frame sources from the brand's own in-play, adjacent, and out-of-play visual language rather than the model's default guess. The Parker tools that pull brand reference and creative data are inventoried in parker-system/system/parker-tools.md.
Two pieces of context every AI-generation prompt depends on:
Identify the prompt type. Run strategy.md to pick which process. Video or static, and within that, which variant — single-shot, multi-shot, image-driven, dialogue-driven, format-generation, or recreation.
Load brand context — strategy first. The brand's committed strategy (strategy/ — the working thesis, the roadmap's calls on what to say and show) frames what's worth generating; read it before writing the prompt, and check the idea bank (idea-bank/, including evaluated ideas) for an entry this asset should execute from. Then brand voice, compliance, ICP, visual identity, current creative challenges. The prompt must be grounded in this. A request that cuts against the committed strategy is surfaced with the conflict named, not silently executed; a brain without strategy/ or an idea bank yet gets one line saying so. AI-generated assets that ignore brand voice produce technically-correct output that is functionally useless because it does not look or sound like the brand. When sub-context-docs/visual-vocabulary.md exists, load it too — the brand's in-play filmed settings, talent situations, and product moves are the visual language the generation prompt grounds in, per the method at parker-system/creative-strategy-context/visual-vocabulary-method.md.
Apply the format library or the five-part formula. Video uses the five-part formula — cinematography + subject + action + context + style. Static uses the format library — proven prompt structures with bracketed fields for brand-specific content.
Direct audio explicitly for video when sound matters. Dialogue in quotes with delivery direction. SFX as named sounds. Ambient soundscape. Music register.
Source claims from verified data. Every stat, every claim, every customer quote must come from brand context, customer reviews pulled via tools, ad comments, or data the user explicitly provided. If a verified source does not exist, mark the gap with [STAT NEEDED — verify before publishing] and leave it.
Run the final checklist.
You run the two review gates yourself, now, before you present anything — this is a step you execute, not a thing that happens for you. Read "automatically" as without being asked, never as without you doing it: nothing else spawns these agents, so if you don't spawn them they never run. They are not optional, not a "second opinion," and never offered to the user as a choice — asking "want me to review this" is itself the failure. Running scripts/voice-lint.py yourself is not the gate; the gate is the independent agent, and its returned verdict fills the receipt block below, which you cannot write yourself. Grounding gate first (it changes content), voice gate second (it changes lines). Spawn the context-grounding-review agent (defined at .claude/agents/context-grounding-review.md) with the user's task, the prompt output, the brain root, and the list of tool pulls made this session. It runs scripts/grounding-check.py and independently derives what should have been loaded and pulled — the visual vocabulary behind the frame, the verified sources behind every stat and quote, the brand reference behind the visual identity — then diffs that against the evidence. A bounced verdict means re-pull and regenerate the affected parts. This gate runs before the voice gate because its verdict changes content, not lines. And every bounce gets captured through self-improvement-intake as a one-line reasoning trace — the task shape plus the loads or pulls that were missed — so the routing layer learns from the catch instead of re-making the same mistake for the gate to re-catch.
Run the voice review gate on the customer-facing copy. The words a viewer will read or hear — dialogue lines in quotes, headlines, on-screen text, supporting copy, recreated copy — pass the gate; camera and technical direction do not. Spawn the creative-voice-review agent (defined at .claude/agents/creative-voice-review.md) with those lines extracted, the brand voice profile if one exists, and the deliverable type. The agent runs the mechanical lint (scripts/voice-lint.py) and the judgment pass per ai-writing-tells.md — and spoken-script-voice.md for spoken dialogue — independently, in a context that did not write them. Apply its rewrites into the prompt and re-run until the verdict is ships. A flag that conflicts with a sourced customer quote keeps the quote, with the reason carried into the Voice Review block.
Format output per the structure below.
SHOT BRIEF — Shot type, camera movement, subject, core action, mood/style, audio elements, aspect ratio, duration. THE PROMPT — The complete, self-contained text the user pastes into Veo. WHY THIS PROMPT WORKS — Two-to-four-sentence rationale on the cinematography, the action specificity, and the audio direction.
FORMAT BRIEF — Format name (from the category library), what it tests, aspect ratio. THE PROMPT — The complete prompt the user pastes into the image model, with brand-specific values filled into all bracketed fields. WHY THIS FORMAT FITS — Two-to-four-sentence rationale on why this format matches the brand, the ICP, the current moment.
THE STORY THIS AD IS TELLING — Full paragraph explaining the original's narrative — what it wants the viewer to believe, how the visuals and copy work together to build conviction, why it works. RECREATION BRIEF — for each section that applies — headline, supporting copy, design — give the original and then the brand's version, written as prose, not a table: name the section, quote the original, then the rewrite, in plain sentences. Skip the sections that do not apply. COMPLIANCE FLAGS — Anything in the recreation that requires a disclaimer, a forbidden term, or stat verification.
Append to every output:
The grounding gate's receipt, required on every prompt: the context-grounding-review verdict, one or two plain sentences on what it checked (visual vocabulary evidenced, stats and quotes traced, cited sources resolved), and — if it bounced — what was re-pulled and regenerated before re-shipping. A prompt missing this block did not pass the gate.
The gate's receipt, required whenever the prompt carries customer-facing copy — dialogue, headlines, on-screen text: the creative-voice-review verdict, the lint density before and after, one or two plain sentences on what was flagged and fixed, and any flag kept with its reason. A prompt whose copy skipped this block did not pass the gate. Prompts with no customer-facing copy at all state that in one line instead.
sub-context-docs/visual-vocabulary.md exists, the brand's in-play filmed settings, talent situations, and product moves ground the generation prompt — that is the brand's visual language. Any visual invention outside the vocabulary is out-of-play; flag it so the user knows the prompt is asking for something the brand has not shot.[STAT NEEDED — verify before publishing] and leave the gap.context-grounding-review agent verifies the prompt was built from the brand's visual vocabulary, verified data, and traced customer language — an invented quote or an unresolved cited source is an automatic bounce. Its verdict appears in the Grounding Review block.creative-voice-review pass. The independent agent runs the mechanical lint (scripts/voice-lint.py) and the judgment pass against ai-writing-tells.md on every line a viewer will read or hear; its verdict and the before/after lint density appear in the Voice Review block. Self-review does not substitute — the reviewer must not be the writer.