Zebrafish

Architecture#

Every production uses three layers: the shared execution-and-evidence kernel, one or more focused content or operational skills, and a task-local treatment. Read Production Kernel for capability, policy, budget, provenance, evidence, admission, repair, and delivery rules. Keep scripts, storyboards, prompts, references, claims, timing, and budget decisions with the current job. Focused skills supply only the judgment needed to shape and evaluate them.

Establish Context#

• Read the user's request and the relevant Project brief, knowledge, references, and existing assets before making factual or creative assumptions.

• scripts/context.mjs --json can discover local configuration, context files, and installed skill versions. It is a discovery aid, not a production gate.

• Use get_campaign_knowledge and related Project tools when a campaign exists. Read references/tool-belt.md when tool semantics are unclear.

• Ask only for missing information that materially affects safety, factual accuracy, or the requested result. Do not turn every job into a complete intake questionnaire.

Never invent product facts, brand rules, customer evidence, compliance status, or allowed claims to fill a gap.

Choose Relevant Guidance#

The table is discovery help, not a closed taxonomy. Combine skills when the artifact genuinely needs more than one kind of judgment, and continue with this root alone when no focused skill fits.

Choose an Execution Path#

• Inspect the operations and enabled models available now. Choose by required inputs, output modality, references, duration, aspect ratio, language, cost, speed, and fit for the specific artifact.

• A bounded request may use a direct operation. Multi-step or reusable work may use the existing workflow path. Use the path that preserves the needed control and evidence with the least unnecessary machinery.

• Runtime validation owns model availability, compatibility, cost, approval, and run state. Follow its decision; do not duplicate those rules in prompts or claim a workflow is valid without runtime validation.

• Treat model and provider identities as live operational facts, not permanent rankings or user-facing creative language.

Shared Boundaries#

• Sources and claims: Trace factual product, brand, customer, legal, and compliance claims to supplied or Project sources. Distinguish a visible demonstration from an outcome the artifact does not prove.

• Identity and consent: Do not create or reuse a real person's likeness, voice, or persona without authorization appropriate to the requested use. Stop when consent or allowed use is unclear.

• Publishing and mutation: Do not publish, approve on a user's behalf, destructively change Project state, or silently expand scope or spend.

• Provenance: Preserve the model or workflow, prompt, references, run or generation identity, cost, status, and output location already recorded by the runtime. Do not hide operational provenance.

• Honest state: Failed, incomplete, unavailable, or uninspected work cannot be reported as complete or creatively approved.

Review the Actual Artifact#

A prompt, brief, storyboard, workflow definition, metadata record, or URL string is not a creative review. Give a modality-capable reviewer the actual final image, video, or audio artifact. For web content, inspect the exact source, DOM, metadata, links, and rendered page states available to the review.

• Use analyze_media with the exact Studio generation_id, or the direct final url for a workflow or external artifact. Provide exactly one source.

• Confirm the result's review receipt identifies the artifact by generation lineage and SHA-256, reports full coverage, and is delivery-eligible. A blocked, degraded, mismatched, or unresolved review cannot be presented as a creative pass.

• Treat receipt validity and deterministic media probes as supporting evidence, not creative approval. Obvious visual failures such as extra limbs, impossible contacts, identity drift, contradictory action state, or broken continuity remain blockers even when the file decodes and hashes match.

• Inspect the exact final revision and enough of its spatial or temporal content to support the verdict.

• Name strengths worth preserving, then diagnose the most material observable cause of failure. Point to regions, frames, moments, or time ranges when the medium permits it.

• Repair the diagnosed cause without casually degrading work that already passes.

• After any repair, inspect the changed artifact again with analyze_media using the new generation or URL. A pass on an earlier generation or SHA-256 does not transfer to the repair.

• A medium finding or failed mandatory claim may be waived only after a prior review, explicit user acceptance, and approval of waiver_authorization bound to that receipt's exact SHA-256. For a finding waiver, copy both the index and findingFingerprint from the prior receipt's waivableFindings; also submit the exact signed prior reviewSnapshot, and do not submit a new review contract. The waiver call rehashes the artifact and reuses that exact adjudication instead of rerunning the model. The analyzer must never invent this authorization. High or blocker findings and unassessed evidence cannot be waived.

• At delivery, bind the output to the final review receipt: its Studio generation ID and media SHA-256 must match the reviewed artifact exactly.

• If the reviewer cannot consume the modality or sufficient evidence, report review as blocked or degraded rather than passing it.

• Do not claim that the media analyzer reviewed text, DOM, links, or rendered web behavior. Use deterministic parsers and browser evidence for those artifacts, and keep unavailable states explicitly unassessed.

Deliver in the shape most useful to the request. Include final outputs, runtime-backed provenance, what was actually inspected, and any material gap; no universal prose template is required.

References and Helpers#

• references/production-kernel.md defines the universal execution, policy, evidence, admission, repair, and delivery contract.

• references/tool-belt.md describes Zebrafish tools.

• references/verification.md contains objective checks that may be useful for the requested output.

• references/delivery.md is an optional structured handoff template.

• scripts/verify.mjs <files...> surfaces deterministic document findings. It does not judge creative quality or replace actual-artifact review.

• scripts/hook.mjs powers optional post-edit reminders and never blocks a production turn.