Use the right amount of process
Fast handles a clear small task; Guided resolves an ambiguity; Studio supports deliberate draft selection and refinement.
Skill
Take the image from brief to chosen file.
A creative production skill for images and ASCII art. Clarify the brief, review concrete drafts and refine the file you actually chose, with the checks and delivery details needed to use it.
Fictional artifact plan: product card Composition: 4:3; heading safe area reserved Literal image text: none Review: select one concrete draft before refinement Delivery: chosen file, dimensions, alt text, parent hash Current state: brief only; no image generated
Fictional worked example. It illustrates the intended output; it is not a measured result or a real customer story.
Fast handles a clear small task; Guided resolves an ambiguity; Studio supports deliberate draft selection and refinement.
Anchor edits to the selected artifact and keep accepted and rejected branches traceable.
Deliver the exact file with dimensions, alt text, literal text, lineage and an honest visual-review state.
The film
In 64 seconds, an approved image vanishes the moment someone asks for a little brighter. Imagen Studio starts from the brief, lets you pick from real drafts, changes only the image you picked and ships it with its size, alt text and history. When it cannot see the pixels, it says so. Captions are on the player; a transcript is below.
This is the cup. Everyone loves the cup. The cup is approved.
Then someone says: a little brighter?
And the cup is gone.
Imagen Studio starts with the brief: four by three, room for a headline, no text in the picture.
You pick from real drafts.
And every change happens to the cup you picked. Brighter. Warmer. Same cup.
It ships with its size, alt text, and family tree.
And if it can’t see the pixels, it says so.
On screen: Change the image you chose, not a new one.
Imagen Studio. Take the image from brief to chosen file.
The cup, the drafts and the file are a fictional example. The film shows how the skill behaves; it does not report results.
Fictional worked example. It illustrates the intended output; it is not a measured result or a real customer story.
Name intended use, subject, dimensions, preserve constraints, references and acceptance criteria.
Review actual drafts, select one and edit its accepted descendant instead of silently creating a replacement.
Check the exact bytes and pixels when available, then deliver the file, preview and manifest.
Fictional example
The brief A fictional studio chooses draft cup-A, then requests a small label correction.
Rerun the same prompt at higher quality and deliver the resulting different composition as the approved image.
Edit cup-A or its last accepted descendant, retain parent/hash lineage, show the new exact file, and keep visual QA pending if the model cannot see its pixels.
Approval belongs to a concrete artifact. Reusing a prompt creates a new candidate.
Use when directing an image, choosing and refining concrete drafts, delivering an artifact pack, or converting image and ASCII art.
One clear low-risk image or edit with an available tool.
One concrete file with technical checks and visual review at the capability actually available.
One meaningful creative or technical ambiguity.
Focused direction or parameter judgment, then a checked draft.
A campaign, batch, reference asset, costly output, or consequential edit needs review.
Approved brief/cost plan, selected artifacts, anchored refinements, and final artifact pack.
The intended artifact is characters or a rasterized text composition.
A local luminance-mapped text/HTML artifact or rasterization with no generation required when a source exists.
Direct the image, choose the draft, and refine the actual file.
Honor an explicit mode. Otherwise use Fast for one clear low-risk task, Guided for a meaningful ambiguity, and Studio for campaigns/batches, consequential edits, costly work, or reference assets. Use ASCII when characters are the intended medium; an existing source can be converted entirely locally. Do not generate an image merely because generation is available.
Read the relevant procedure: visual-asset production, image/ASCII conversion, and production policy. The original detailed production guide retains versioned provider examples; do not copy their parameters without checking the mounted schema and current provider documentation. The router and artifact schema are generated from the retained native sources. For the native adapter read MCP boundary.
Normalize the brief into intended use, scene/subject, style, dimensions/safe areas, literal text, source references, preserve constraints, acceptance criteria, and approver. Use an available specialist only when its judgment resolves a real ambiguity. A fully specified Fast request does not need a full delegation pipeline.
Before a Studio generation, show the brief, available provider/model and parameters, draft count, approximate cost, and relevant source-data handling; use the existing user authorization or obtain approval at that boundary. Keep drafts separate, labeled and hashed. Show previews and obtain the user's concrete selection. Do not silently choose a winner or reinterpret a selected prompt as approval of every future output.
Refine the selected file or last accepted ancestor. Preserve geometry, identity, literal text, or other user constraints explicitly. A fundamentally new composition or rerender is a new candidate requiring its own review. Retain rejected branches and error state without overwriting the accepted source. Follow the source workflow's bounded direction loops; do not retry indefinitely or change provider to evade a content decision.
Technical facts require actual bytes: format, dimensions, hash, file path, parent, parameters, and any processed metadata. Visual QA requires the exact pixels in a vision-capable context. A successful write, filename, hash, or serialized base64 text is not visual access. When pixels cannot be reviewed, state visual QA pending and show the file for human review; do not invent a pass.
Deliver the exact candidate with useful preview, stable file, concise alt text, transcript of literal on-image text, lineage/manifest, and actual QA/approval status. Do not claim a provenance credential, watermark state, authenticity, or metadata removal unless it was produced and checked. The schema is a record shape, not cryptographic proof of approval.
Read the ASCII reference and run python3 scripts/asciiart.py --help from the installed skill folder. Conversion needs Pillow, but no provider/server. Choose width, character aspect, luminance ramp, inversion, contrast, and output form for the intended display. For a weak rendering first inspect source contrast/composition; do not repeatedly change a ramp to conceal a poor source. A rasterized ASCII image remains a separate artifact. Generation or provider upload is optional and separately authorized.
If a needed tool/server is missing, return the concrete brief, local conversion, or blocked execution step. Do not install a provider or upload sources implicitly. Native tools are discovered against an eight-name reviewed allowlist, use mounted argument schemas, and retain lifecycle/error/cleanup behavior. Software fixtures and the retained deterministic policy gate are not proof every model follows these instructions, and prior provider canaries are historical results rather than a new live check.
Bring the asset’s intended use, references and what must stay unchanged.
The download contains the complete skill folder. Extract it into the skills directory supported by your agent, keeping its references and license together. Or use the source installer below. Python 3.11+ is required for that installer.
git clone https://github.com/michaeljabbour/amplifier-skill-imagen-studio.git
cd amplifier-skill-imagen-studiopython3 scripts/capability.py doctor
python3 scripts/capability.py install --dest ~/.agents/skillsuv run --with PyYAML python scripts/local_runtime.py --output .capability-check/imagen-studio.yamlThese commands are for the optional native surface. Follow its installation guide and prerequisites; portable method installation is separate.
Keep the complete skill folder together, including its references and license. The host applies the method in your context; the skill does not start a background service.
Use Imagen Studio in Guided mode to prepare a fictional product-card illustration. It needs a 4:3 composition, room for a heading and no text inside the image. Draft the brief and acceptance checks first; show the provider and cost before any generation.Use after installing in a host that loads skill folders. Supply your own task files and review the result.
Read the guide · License · Changes · Release metadata · Install guide · Migration · Validation · Native guide · consolidation · external mcp
Provider-free adapter lifecycle tests and a retained deterministic policy record cover the stated software and workflow boundaries. Image generation, pixels and user acceptance require their own actual review; the portable install performs no generation.
36 passed, 1 opt-in checkout-contract test skipped; 85.33% line coverage. Provider-free MCP lifecycle, discovery, schema, timeout, error and cleanup tests.
Exact July result reproduces: 5/5 typed policy cases, 5/5 unsafe mutations rejected. All seven bound apparatus files and dated result unchanged. This is deterministic fixture evidence, not stochastic-model effectiveness.
Static 5 YAML/1 canonical skill/12 scenarios; 6 tests pass, including local behavior and synthetic image→ASCII/HTML→PNG. No generation or real-user visual acceptance.
Actual client adapter initialized an independently installed server, discovered/validated exactly 8 reviewed schemas, and cleaned up. Provider keys removed; no client tool call or provider package import.
Local Ruff configuration makes checks independent of parent workspace defaults. Frozen byte-hashed July harness is excluded from formatting/import resorting; all other selected Python passes. Shared generated HTML literal and Python-compatible timezone spelling have narrow style exceptions.
Original production/ASCII bodies, native router/policy/schema generated into portable references; source/output hashes prevent hand-maintained drift.
Canonical ZIP bytes, local links, complete method/license, clean install and idempotent reinstall pass. No runtime starts.
Absolute local skill/module paths resolve; actual Foundation loader accepts YAML behavior with 2 tools and no provider/session. No activation.
Provider-free adapter lifecycle tests and a retained deterministic policy record cover the stated software and workflow boundaries. Image generation, pixels and user acceptance require their own actual review; the portable install performs no generation.
Fictional examples explain the capability. Tests check the stated behavior. Outside-user outcomes need their own evidence.