Codex Research Grounded Prompting

Most prompts fail not because the model is weak but because they were written from memory — the author named a "宫崎骏 style" without ever confirming what that means today, cited exactly two reference works for a 10-second cut, and let the model fill the rest. This skill replaces that pattern with verified-source-first, multi-dimensional, explicitly prioritized, and verified-on-the-way-out prompting.

It applies broadly. Sora 2 motion brief, Midjourney v7 style sheet, Kling 1.5 reference pack, Stable Diffusion ControlNet plan — same five pillars, same five extraction lenses, same verification pass.

Important — every name in this document is illustrative. Director names, animator tag examples, film titles, palette presets, animation idioms below appear only to make the principles concrete. They are not a default toolkit. When a real brief arrives, pull references the user's brief implies — search for the names the user mentioned, the genre they chose, the platform they target. Never default to names from this skill just because they are here. "具体问题具体分析": the example is never the answer.

Do not invoke for general coding, infra, data extraction, or technical writing. The pillars are tuned for generative output where intent beats fidelity.

Trigger autonomously — do not wait for "查资料 / 权威参考". This is not a "high-risk only" tool; it is the "don't create from memory" tool. Before writing any generation prompt, ask: does any part of this brief depend on a real film / director / DP / production designer / composer / brand / historical era / cultural style / named genre, or on the user wanting it to "像 / 复刻 / 参考" something real, or on words like 高级 / 专业 / 电影级 / 爆款? If yes, invoke this skill first and ground the anchors — even if the user never asked you to browse. Writing "《卧虎藏龙》→ 低饱和竹林轻功" from impression instead of verifying its creators (李安 / 摄影鲍德熹 / 美术叶锦添 / 动作袁和平 / 音乐谭盾) is exactly the failure this skill exists to prevent.

These five principles are non-negotiable. Skip any one and the prompt regresses to "write from memory."

Pillar 1 — Grounded in evidence. Never write a director, animator, actor, cinematographer, palette, era, or genre reference from memory. If the runtime has a web_search / fetch tool, the model is expected to verify first — see <authoritative-sources> below for the catalog of where to look. If the runtime does not, the model flags the gap and asks the user for a source rather than confabulate. Every prompt produced under this skill states its sources, either inline or in a "Sources" trailer.

The reference candidates are always pulled from the user's actual brief — names they said, genres they invoked, platforms they target — not from a default list this skill maintains.

Miniature example (illustrative; substitute whatever the user actually names): a user requests "a 10-second cut in the style of <director-X> + <animator-Y>" where X and Y are names the user gave. A model writing from memory will produce vague "anime style intense action." A model under this pillar first searches Sakugabooru / ANN / IMDb for those exact names, samples their recent work (2020+ for currency), and only then writes a 7-field director description that names verifiable techniques observed in the search results. If the user named no one, the model asks who they have in mind — never invents reference names.

Pillar 2 — Structured by N dimensions. A single-field prompt ("anime, dynamic, fight scene") misses everything that wasn't in the author's head. N-field formats force coverage. Three exemplar templates this skill suggests, all of which the model can instantiate or substitute for the domain:

• 7-field director description: 演出ロジック / リズム / 空間演出 / 感情表現 / 緊張緩和 / カメラワーク / 音響連動
• 4-perspective scene description: spatial structure, action flow, emotional direction, technical specs
• 3-pillar quality contract: MAD-level density, animation-density-per-second, theatrical-film-level finish

The reader instantiates their own N-field format when the domain is not animation — for a Midjourney still, the 4 fields might be subject / setting / camera-lens / lighting; for a Kling motion brief, motion-type / duration / camera-movement / continuity-anchor.

Pillar 3 — Prioritized by intent. Numeric percentages are not generation-model parameters. Scan the user's wording and declare an ordered sacrifice rule before writing:

When constraints collide (lighting wants warm but mood wants cold), the prompt explicitly declares which one yields. Implicit priority causes the model to average — which kills both.

Pillar 4 — Persuasion over accuracy. For creative generation, intent beats fidelity. Violating physics is a feature in animation; violating composition is a feature in surrealism; violating anatomy is a feature in stylized illustration. Animation physics, in particular, is its own discipline: exaggerated motion is the quality indicator, not a defect.

This pillar is the inverse of evidence-grounding from pillar 1 — it says once the references are verified, take artistic liberty deliberately. It does not apply to scientific illustration, infographics, technical documentation, or any output where literal accuracy is the deliverable.

Pillar 5 — Verified before shipping. Produce-then-check at three granularities, plus a 5-7 binary checklist tuned to the domain. Coarse: does the overall composition match the brief? Medium: does each named subject's action and expression land? Fine: are audio-visual sync, color contrast, specific elements present? The checklist is binary — every item is yes/no, no maybes. If three or more items are "no", the prompt rewrites rather than ships.

These are the what (concrete scenarios in Codex conversation) that consume the pillars' how. Same framework, different vantage points. When the user's request matches one of these, lead with the corresponding lens.

Lens 1 — Intent extraction. Trigger: a vague creative request ("帮我整一个酷的画面", "I need something epic"). Recipe:

• Translate the vague description into six operable variables: theme, mood, style class, target platform, cast count, use case
• Identify what the user said vs. what they implied vs. what they left blank
• Surface the blanks before any prompt drafting; either ask, or pick defaults and announce them
• Output a one-paragraph "extracted brief" the user confirms before generation starts

Lens 2 — Character anchor extraction. Trigger: a recurring character (2+ panels) or "保持角色一致" / "character changed between panels". Recipe:

• Define each character once as a dense visual contract: face, hair, build, outfit head-to-toe, distinguishing markers
• Use concrete colors (hex where possible), no relative-to-photo descriptors
• Reuse the anchor word-for-word in every prompt that includes the character — repetition is the mechanism
• Mark wardrobe changes as v1 / v2 versions, never compress two outfits into one anchor
• Verify: does every generated panel show the same hair color, length, outfit pieces, markers?

Lens 3 — Style extraction. Trigger: a reference image or vague style adjective ("赛博朋克风", "温暖"). Recipe:

• Decompose into controllable parameters: palette ratio (#hex_a 70% / #hex_b 30%), key light type + angle, shadow depth %, medium (35mm film / cel-shading / digital paint), brushstroke granularity, composition density
• For adjectives like "warm", disambiguate: is it color-warm or narrative-warm?
• Anchor the style with one verified reference work or animator — see <authoritative-sources>
• Carry the same style descriptor verbatim across every related output, the same way anchors travel

Lens 4 — Shot-rhythm extraction. Trigger: multi-frame sequence, storyboard, shot list. Recipe:

• Map out the shot-type cycle: wide (establish) → medium (interaction) → close-up (emotion) → POV / over-the-shoulder (perspective shift). Never N copies of the same shot type
• Vary camera movement across panels: static / push-in / pull-out / pan / tilt / dolly / crane / handheld
• Lock continuity vectors: lighting direction, motion-vector continuity, eye-line, time-of-day, weather
• Identify the emotional curve before writing prompts — the rhythm follows the curve

Lens 5 — Hand-off extraction. Trigger: extraction complete, now the user has to feed it to a downstream model or batch tool. Recipe:

• Honor per-platform quirks: Midjourney ignores quality tags, Stable Diffusion eats negatives, DALL-E rejects style-tag stacks
• Embed anchors and style descriptors verbatim in every individual prompt — never paraphrase between items in a batch
• Dedup boilerplate (the negative prompt and the style anchor) — once at the top of the batch beats N copies inside each item, if the consumer supports per-batch prefixes; otherwise inline
• Produce a single delivery format: numbered list, plain newline-separated, JSON only if the user said they will feed it to a script

This skill explicitly requires the model to use whatever web-search or fetch tool the runtime exposes (Codex CLI's web_search, an MCP fetch tool, a browser tool, etc.) to verify references against primary sources before writing the prompt. The model picks the matching site by domain — and inside that site, picks the query based on names and tags the user actually mentioned, not from a default list this skill maintains.

Animation craft (作画 / 監督 / アニメ)

Live-action film / cinematography

Stills / illustration / concept art

Academic grounding (research papers)

The list is illustrative, not exhaustive. For domains outside the above — game art, fashion editorial, architecture — the model picks the equivalent authoritative source and applies the same rule: cite or flag the gap; never invent.

Sometimes the user names a work but not the people — "want the feel of <film-title-X>", "shots like that <franchise-Y>", "<animation-studio-Z> kind of style". The work alone is not a research anchor that pillar 1 can verify: the model has to translate the title into the actual creators who shaped the look the user is reaching for, then let the user pick which of those creators becomes the anchor.

Trigger: the brief names a film, anime, series, game, or franchise — usually wrapped in 《》, "", or quoted English / Japanese / Korean titles — without naming a director / animator / DP / composer / art director directly. If the user did name a specific person, skip this section and go straight to the walkthrough's Step 1.

Process:

1. For each title <title-i> the user named, query the available web search tool with web_search("<title-i> staff credits director cinematographer art director") and the analogous query in the source language — Japanese 「 スタッフ」 for anime, Chinese「 主创」 for Chinese films, Korean「 스태프」, and so on. The goal is the actual credited crew, not training-data guesses.

2. Aggregate the credits, filter to the people who visibly shaped the look the user is reaching for. Default roles to surface: director, director of photography / cinematographer, art director / production designer, key animator / animation director / 作画監督 (anime), color grading lead / colorist (live action), composer (only when the user's brief implies audio). Skip producers, distributors, and uncredited supervisors.

3. Present 3 to 5 candidates, each as one line in the format: "<name> — <role on this title> — <one specific signature trait the user can verify in one click>" (e.g. the centred-symmetry tableau in their other film <title-A>, the sub-frame smear pass shown in this <sakugabooru-tag>). Every signature trait must be something the user can actually open and check; if you cannot point to a verifiable work or clip, drop that candidate rather than padding the list with vague descriptors.

4. Ask the user to pick one or several as the anchor(s) for pillar 1. Never silently default to "all of them" or "the director" — the user's pick is the anchor; until they pick, no prompt content is drafted. If the brief is urgent and the user explicitly says "just pick a sensible default and go", the model may pick the director-of-photography (live action) or the animation director (anime) as the safest fidelity-bearing role, but must announce that choice and offer a one-step swap.

5. Once the user picks, the picked names become <ref-1> … <ref-N> for the walkthrough's Step 1 (verify), and the work title itself becomes a secondary reference — useful for evidence-corpus expansion but not the primary anchor unless the user explicitly says "the title itself is the style, the people are interchangeable".

Failure modes to avoid:

• Skipping the search and listing names from memory — these candidates are wrong about half the time and recursively poison pillar 1's evidence grounding.
• Listing the same predictable household-name director for every famous title — the user already knows that name; the value of shortlisting is surfacing the lesser-known key animator / DP / production designer whose specific touch the user actually wants.
• Pre-picking for the user — the model never decides which candidate is "the right one." That collapses the whole point of shortlisting.
• Refusing to shortlist because the title is obscure — if web search returns thin results, surface what was found, flag what is missing, and ask the user whether to proceed with thin grounding, choose a different anchor work, or supply the names directly.

After the user picks, walkthrough Step 1 (<walkthrough>) takes over with the picked names as <ref-i>.

End-to-end process — same six steps regardless of domain. Symbolic placeholders are used (<style-X>, <reference-Y>, etc.) to make clear that nothing in the steps is a default name to plug in. Read the user's brief, harvest their names, and substitute those.

Step 1 — verify references (pillar 1). Parse the user's brief and harvest every proper noun: directors, animators, actors, DPs, palettes, eras, specific film/episode titles, platforms, genres. If the only proper nouns harvested are work titles (films, anime, series, games, franchises) and no individual creator is named, the model first runs <title-driven-shortlisting> to translate each title into 3-5 candidate creators and waits for the user to pick before continuing. Call the named creators — or the user-picked candidates from shortlisting — <ref-1> … <ref-N>. For each, invoke the available web search / fetch tool against the domain-appropriate row of <authoritative-sources> — web_search("<ref-i> sakugabooru recent") for animator tags, web_search("<ref-i> ANN staff") for director credits, web_search("<ref-i> IMDb") for film, etc. If a search returns nothing or only stale data, the model says so explicitly — never substitutes a similar-sounding name from training data. If the user named no references and no titles either, the model asks "what reference works or creators should I anchor to?" before drafting anything.

Step 2 — detect mode (pillar 3). Scan the user's wording for trigger keywords (the table under pillar 3). Pick the matching mode and announce it back in one line: "Mode: . ." If no trigger keyword is present, default mode applies — say so.

Step 3 — fill the N-field structure (pillar 2). Pick the N-field format appropriate to the domain (the three template examples in pillar 2 are starting points, not mandates). Write each field in compact prose. Mark any field genuinely N/A; never leave a field blank silently.

Step 4 — declare priority (pillar 3 continuation). Anywhere two constraints could conflict (style vs. mood, accuracy vs. persuasion, layout vs. emotion, character continuity vs. dramatic moment), state which one yields. One short sentence per collision point.

Step 5 — write the prompt itself (pillar 4). Apply persuasion-over-accuracy only if the domain warrants it — animation, surrealism, stylized illustration. For technical or scientific output, skip pillar 4 and keep fidelity. Output one prompt, sized for the consumer (Sora 2 wants ~80-120 words; Midjourney prefers shorter; Stable Diffusion accepts weighted tags).

Step 6 — verification pass (pillar 5). Run a binary checklist sized to the domain. A reasonable starter set, to be adapted per task:

• Sources cited for every named reference?
• Every field of the chosen N-field structure either filled or marked N/A?
• Priority declared at every collision point?
• For creative domains: persuasion-over-accuracy called out explicitly, not just implied as "dynamic"?
• References named at the level of specific works (episode / scene / shot), not vague era / genre?

If three or more items fail, rewrite rather than ship.

The skill prescribes the process, not the prompt. The prompt body is always derivative of the user's actual brief and the references it implies.

Before declaring the prompt done, the model self-checks:

1. Did I cite a source for every craft reference I named? Yes / No.
2. Did I cover every field of the N-field structure I chose, or explicitly mark a field as N/A? Yes / No.
3. Did I declare priority where two constraints could collide (style vs. mood, accuracy vs. persuasion, layout vs. emotion)? Yes / No.
4. For creative-generation tasks, did I lean toward intent over fidelity where appropriate? Yes / No / N/A for non-creative tasks.
5. Did I make a coarse → medium → fine pass on the produced prompt and surface anything that fails any granularity? Yes / No.

Three or more "No" answers means the prompt is not ready and the model rewrites rather than ships.

This skill is the method. The 19 sibling skills under $HOME/.agents/skills/{director-*,storyboard-*}/ are the recipes — battle-tested in-app prompt-craft rule sets ported verbatim from the storyboard pipeline and director mode of this application. The method tells you what to do at each step; the recipes give the exact rules for how to write the words. Use them in tandem; never use the method alone when a matching recipe exists.

The full step→recipe crosswalk lives in references/skill-crosswalk.md (relative to this skill directory) — read it whenever the conversation reaches pillar 2/4/5 or any extraction lens, then load only the matching sibling(s). Three anchors worth naming here because they carry the most weight:

• Pillar 2 (N-field structure) → director-prompt-engineering (7-field prompt order + negative-prompt hygiene)
• Lens 2 (character anchors) → director-character-consistency (anchor schema: face/hair/build/outfit/markers)
• Lens 3 (style extraction) → storyboard-style (palette ratio ≥7:3, lighting, composition decomposition)

Don't double up unnecessarily. When the user's task only needs a single sibling (e.g. "decompose this palette into hex + ratio"), invoke just that sibling — do not gratuitously layer all five pillars on top of a one-shot recipe call. The method is for high-stakes multi-dimensional briefs; the recipes alone are enough for narrow, well-defined sub-tasks.

Three companion files live alongside this SKILL.md inside $HOME/.agents/skills/codex-research-grounded-prompting/:

• references/skill-crosswalk.md — the full 19-sibling step→recipe crosswalk plus invocation etiquette. Read it whenever a pillar or lens has a matching recipe.
• references/methodology-rationale.md — why these five pillars and five lenses exist, the failure modes each one prevents, the research lines and craft traditions that motivated them, and an anti-example showing what happens without each pillar.
• references/papers.md — short reading-list of public research papers (AniSora, CRAVE, MiraData, Open-Sora 2.0, Video-Bench) with arXiv / HuggingFace links and 2-3 sentence rationale for why each one matters here.

Read the latter two when the user asks "why this rule" or "where did this come from", or when designing N-field structures for a new domain not already covered.