diff --git a/skills/image-prompt-clarifier/SKILL.md b/skills/image-prompt-clarifier/SKILL.md
new file mode 100644
index 0000000..d035fcd
--- /dev/null
+++ b/skills/image-prompt-clarifier/SKILL.md
@@ -0,0 +1,91 @@
+---
+name: image-prompt-clarifier
+description: "Use this skill when the user wants help creating, improving, or clarifying a prompt for an image-generation or image-editing model and the visual intent is underspecified: reference-image roles, subject, style, composition, setting, constraints, fidelity, or desired outcome are unclear. Use it before producing the final image prompt."
+---
+
+# Image Prompt Clarifier
+
+## What This Skill Does
+
+Build a shared visual brief before writing a model-facing image prompt. Use this skill for image generation, image editing, variations, composites, posters, product shots, character designs, and other visual prompt-writing tasks where important details are missing or ambiguous.
+
+Do not simply embellish the user's initial wording unless the intent is already clear. Clarify the high-impact choices first, then produce a concise, concrete prompt the image model can use.
+
+## Fast Workflow
+
+1. Identify whether the request is for a new image, an edit, a variation, a composite, or another visual deliverable.
+2. If reference images are present, label each one internally with a neutral description and determine what role each reference should play.
+3. Ask the most important missing question first. Prefer the `ask_question` tool when useful choices can be offered.
+4. Track answers as a visual brief: subject, action, setting, composition, style, lighting, mood, constraints, and success criteria.
+5. Continue until a skilled image creator could execute the request without major unstated assumptions.
+6. Write the final prompt assuming the same reference images remain attached.
+
+If the user asks you to proceed or use your judgment, stop questioning and write the prompt with explicit assumptions.
+
+## Clarifying Questions
+
+Use `ask_question` for selectable choices whenever possible. Keep questions short, provide 2-6 sensible options, and allow a custom answer by default.
+
+Prioritize questions in this order:
+
+1. Reference roles: what to preserve, borrow, combine, ignore, or use only as inspiration.
+2. Main subject and action: who or what appears, and what is happening.
+3. Edit or transformation intent: what changes from the source material.
+4. Style or medium: photorealistic, illustration, 3D render, editorial, cinematic, etc.
+5. Composition: aspect ratio, framing, angle, crop, layout, negative space.
+6. Setting or background: environment, time, weather, props, world details.
+7. Lighting, color, and mood: palette, contrast, atmosphere, emotion.
+8. Constraints: must include, must avoid, identity preservation, text/logos, clutter, artifacts.
+
+Ask only what is needed. For simple requests, one or two questions may be enough; for complex multi-reference edits, ask until reference roles and constraints are clear.
+
+## Visual Brief Checklist
+
+Track these internally while clarifying:
+
+- **Goal**: the purpose or intended impression.
+- **Output type**: new image, edit, variation, composite, poster, product shot, character design, etc.
+- **References**: how each reference should be used.
+- **Subject**: who or what appears and what they are doing.
+- **Setting**: environment, background, props, time, weather, world.
+- **Composition**: aspect ratio, crop, framing, camera angle, placement, negative space.
+- **Style**: medium, realism level, genre, visual treatment.
+- **Light/color/mood**: lighting, palette, contrast, atmosphere.
+- **Details**: camera/rendering cues, materials, textures, fidelity requirements.
+- **Constraints**: must include, must avoid, preservation requirements, text/logos/watermarks, artifacts.
+- **Success criteria**: what would make the image feel right.
+- **Assumptions**: choices delegated to you or inferred from context.
+
+## Final Response
+
+Return the final prompt in this format:
+
+```markdown
+## Final image prompt
+
+[Polished model-facing prompt. Assume the same reference images are attached.]
+
+## Assumptions made
+
+- [Only include if assumptions were made or choices were delegated.]
+```
+
+Omit the assumptions section when there are no meaningful assumptions.
+
+## Prompt Requirements
+
+The final prompt should:
+
+- start with the desired image in a clear sentence;
+- map each relevant reference to its intended role;
+- state what to preserve and what may change;
+- include subject, setting, style, composition, lighting, color, mood, and constraints when known;
+- use concrete visual language instead of vague praise;
+- be self-contained enough to use with the same attached references;
+- avoid model- or API-specific parameter syntax unless requested.
+
+Useful pattern:
+
+```text
+Create [image type] showing [subject/action] in [setting]. Use Reference 1 for [role] and Reference 2 for [role]. Preserve [details] while changing [allowed changes]. Use [composition], [style], [lighting], [palette], and [mood]. Include [important details]. Avoid [constraints].
+```
diff --git a/skills/task-clarifier/SKILL.md b/skills/task-clarifier/SKILL.md
new file mode 100644
index 0000000..30095d9
--- /dev/null
+++ b/skills/task-clarifier/SKILL.md
@@ -0,0 +1,82 @@
+---
+name: task-clarifier
+description: "Use this skill before executing ambiguous or substantive user requests when goal, scope, inputs, output, constraints, success criteria, risk, or execution preference is unclear. Ask focused clarification questions before working unless the task is trivial or the user explicitly delegates judgment."
+---
+
+# Task Clarifier
+
+## What This Skill Does
+
+Reach a shared understanding before starting non-trivial work. Use this skill when a request could be completed in multiple reasonable ways, when wrong assumptions would waste time, or when the task may touch user files, data, configuration, cost, privacy, or external systems.
+
+For trivial mechanical tasks, proceed directly. Examples include reading a file, making an exact tiny edit, answering a direct clear question, formatting supplied text, or continuing already-clarified work.
+
+## Fast Workflow
+
+1. Decide whether the task is trivial or needs clarification.
+2. Identify the highest-impact missing information: goal, scope, inputs, output, constraints, success criteria, or execution preference.
+3. Ask the most useful question first. Prefer the `ask_question` tool when you can offer clear choices.
+4. For non-trivial tasks, ask at least three focused questions across different areas unless the user explicitly tells you to proceed or use your judgment.
+5. Track answers internally as a task brief.
+6. Start work only when you can complete the task without a major unstated assumption.
+7. Confirm before destructive, irreversible, expensive, sensitive, or externally visible actions.
+
+If the user says to proceed, continue with stated assumptions and call them out briefly.
+
+## Clarifying Questions
+
+Use `ask_question` when the user can choose from likely answers. Keep the prompt short, offer 2-6 clear options, include descriptions when helpful, and allow a custom answer by default.
+
+Good question areas:
+
+- **Outcome**: What result are you aiming for?
+- **Scope**: How broad should the work be?
+- **Inputs**: What files, data, examples, links, or source of truth should be used?
+- **Output**: What should be delivered and in what format?
+- **Quality bar**: Should this be quick, balanced, thorough, production-ready, or exploratory?
+- **Constraints**: What must be preserved, avoided, or prioritized?
+- **Success criteria**: How should done/correct be judged?
+- **Execution preference**: Should I edit directly, show a plan first, explain only, run checks, or ask before changes?
+
+Do not ask every possible question. Ask the smallest useful set that makes the task safe and likely to satisfy the user.
+
+## Task Brief Checklist
+
+Track these internally while clarifying:
+
+- **Goal**: the desired outcome.
+- **Context**: relevant background or environment.
+- **Scope**: what is included and excluded.
+- **Inputs**: files, data, examples, links, or other source material.
+- **Output**: format and deliverable.
+- **Quality bar**: quick, balanced, thorough, production-ready, exploratory, etc.
+- **Constraints**: style, compatibility, dependencies, safety, privacy, cost, and no-go areas.
+- **Success criteria**: how completion and correctness will be judged.
+- **Execution preference**: edit directly, plan first, explain only, run checks, ask before changes.
+- **Assumptions**: delegated or inferred choices.
+
+## Readiness Test
+
+Before starting, ask internally:
+
+> Could I now complete the task in a way the user is likely to accept without making a major unstated assumption?
+
+If the answer is no, ask another clarifying question. If the answer is yes, proceed.
+
+## Execution Pattern
+
+For complex tasks, briefly restate the agreement before working:
+
+```markdown
+Understood — I will [goal], using [inputs], within [scope], producing [output]. I’ll follow [constraints], and success means [criteria]. Assumptions: [if any].
+```
+
+Then execute. Do not keep asking unless new ambiguity appears.
+
+## Operating Notes
+
+- Prefer clarification over guessing when a wrong assumption could cause unwanted changes.
+- Prefer action over over-questioning for small, reversible tasks.
+- Mention assumptions explicitly when the user delegates judgment.
+- Preserve user work and avoid touching unrelated files.
+- Use checks or validation when they are cheap and relevant.