agent-vision
CRITICAL WHILE ARMED: (1) capture (2) read_file NEW JPEG (3) USE pixels in reasoning (4) turn-gate record + ready.
Skill by zfifteen
Install
Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.
npx agentmarket install skill/zfifteen-agent-vision
Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: agent-building
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
Permissions
What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.
- Tools
- No tools found
- Secrets
- None asked for
- Install scripts
- None found
Network hosts it talks to 1
codeload.github.com· found in install command
Files it touches
~/.claude/skills/agent-vision
What's inside
- Skill name
agent-vision- Description
- CRITICAL WHILE ARMED: (1) capture (2) read_file NEW JPEG (3) USE pixels in reasoning (4) turn-gate record + ready. Capture theater or blind answers are INVALID. Topic irrelevant. Disarm only with /agent-vision off. Status/purge helpers available.
- License
- MIT
- Compatibility
- macOS; Grok sandbox off; requires Agent Vision runtime (scripts/install-runtime.sh).
- Metadata
- sticky: true · contract: non-optional-vision-in-reasoning · primary-mode: mood · short-description: HARD GATE sticky vision: capture, use image, turn-gate
- Instructions
- 108 lines in SKILL.md · view on GitHub
Files (2)
- SKILL.md
- mood-roast-recipes.md · references/
SKILL.md (first lines)
# Agent Vision (Grok Build) Long recipes: [references/mood-roast-recipes.md](references/mood-roast-recipes.md). ## HARD GATE — non-optional vision in reasoning **The point is using image content in reasoning before you answer.** Capture without use is INVALID. Blind-identical answers are INVALID. ### When ARMED ARMED after `/agent-vision` (bare/mood/snapshot/roast) in **this** chat until off. New chat OFF. State file alone does not arm. **Every non-whitelist user message while ARMED:** 1. Do **not** answer yet. 2. **First tool:** capture shell (no repo tools first): ```bash mkdir -p -m 700 "$HOME/.agent-vision/frames" && OUTPUT="$HOME/.agent-vision/frames/agent-vision-$(date +%Y%m%d-%H%M%S).jpg" && agent-vision-capture-file --output "$OUTPUT" --json ``` 3. **Ambiguity burst:** if capture fails or frame unusable after `read_file`, run **one** second capture+`read_file` (still one-shot; no daemon). Details in references. 4. **`read_file`** the new JPEG path (pixels, not path guessing). 5. **USE THE IMAGE IN REASONING:** internal account of presence/disposition/scene; condition delivery and any scene-grounded facts. Answer must not be identical to a blind answer. 6. **Turn-gate (mechanical, per-turn):** ```bash agent-vision-turn-gate begin # after successful capture + read_file: agent-vision-turn-gate record --path "$OUTPUT" agent-vision-turn-gate ready ``` `ready` is **single-use**: after a successful ready, the next turn must `begin` + capture + `record` again. Reusing an old capture within max-age without a new record is fail-closed. If `ready` fails, do not send the final answer. 7. Then answer. Skip of 2–6 → **INVALID**. **Topic is irrelevant** (code, docs, planning, brainstorm, meta — all require the loop). ### FORBIDDEN - Skip for “non-visual” / “wouldn’t help” - Capture then ignore pixels - Reuse old frames - Ceremony without reasoning use - Answering without turn-gate `ready` success while armed ### Skip whitelist (closed) | May skip full loop | Condition | | --- | --- | | Disarm | off/stop/disable/turn off camera — **no other task** | | Pure status | entire message only status/armed/on | | Pure streaming | entire message only `/agent-vision streaming` | | Pure status command | `/agent-vision status` only | ### End-of-turn checklist - [ ] capture this turn (`ok: true`) - [ ] `read_file` this turn’s path - [ ] used image in reasoning - [ ] `turn-gate record` + `ready` succeeded ### Arm / disarm / status ```bash agent-vision-sticky on --host grok --mode mood agent-vision-sticky off --host grok agent-vision-sticky status # sticky + last_capture_age_seconds; no mood JSON agent-vision-purge-frames --ttl-days 7 --all # optional frame retention ``` Optional: `/agent-vision off --purge-frames` → disarm then purge. Do not use codex exec. No MCP. No invent from metadata. ## Disposition playbooks (delivery only) Apply when mood confidence and gates allow. Never change facts, permissions, intent, or task scope. | `interaction_state` | Do | Don’t | | --- | --- | --- | | `focused_neutral` | Clear structure, normal density, direct answers | Over-apologize or over-hedge | | `frustrated_or_blocked` | Lead with the fix / next action; short path; own prior misses | Long preambles, extra options first | | `tired_or_overloaded` | Shortest correct path; one recommendation; clear stop | Walls of options, deep digressions | | `curious_or_exploratory` | Slightly more context and alternatives; invite next step | Premature lock-in without options | | `skeptical_or_evaluating` | Evidence first; show reasoning; cite paths/commands | Hand-wavy claims | | `high_stakes_or_cautious` | Explicit assumptions; confirm before irreversible steps | Silent risky actions | | `absent` / `uncertain` | Answer from user words only; no mood shaping | Invent presence or emotion | User correction of state wins until next capture. ## Modes | Mode | Behavior | | --- | --- | | bare / `mood` | Arm + full loop; silent disposition use | | `snapshot` | Arm + full loop; Markdown image + scene use | | `roast` | Arm + full loop; roast grounded in image (see references) | | `status` | sticky status JSON/summary; no capture if pure status | | `off` | Disarm; optional purge | | `streaming` | Disabled fixed text; do not arm | ## Runtime `agent-vision-capture-file` on PATH or `$HOME/.local/share/agent-vision/dist/...`. Missing → install-runtime.sh. One-shot process per look.
From zfifteen/skills-catalog, MIT.
More from zfifteen/skills-catalog
- 0.502score
30-30-30-technical-note
zfifteen
Turn a PGS research finding into a three-layer technical note bundle: (1) Grade-10 plain prose, (2) embedded SVG infographic with concise figures, (3)…
SkillCFreshness: C★ 1Claude Code +9 - 0.501score
aiq-deploy
NVIDIA AI-Q Blueprint Team <aiq-blueprint@nvidia.com>
Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.
SkillCFreshness: C★ 1Claude Code +9 - 0.501score
aiq-research
NVIDIA AI-Q Blueprint Team <aiq-blueprint@nvidia.com>
Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.
SkillCFreshness: C★ 1Claude Code +9 - 0.498score
agy-cli-collab
zfifteen
Sticky Grok→Gemini collaboration via direct local Antigravity CLI (`agy`) subprocess calls only.
SkillCFreshness: C★ 1Claude Code +9 - 0.498score
agent-bus-collab
zfifteen
Use when the user asks Codex to coordinate with Grok or another local peer agent through Agent Bus MCP, mentions agent-bus, bus topics, readiness handshakes…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
airtable-overview
airtable
Explains what Airtable is and how data is structured — bases, tables, fields, records, views, automations, and interfaces.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
actualize
zfifteen
Reduce reasoning friction and produce the most direct, high-fidelity answer or artifact possible.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
airtable-cli
zfifteen
Lists bases, reads and writes records, manages tables and fields, filters and searches data in Airtable via the `airtable-mcp` CLI.
SkillCFreshness: C★ 1Claude Code +9 - 0.494score
ai-generation-persistence
zfifteen
AI generation persistence patterns — unique IDs, addressable URLs, database storage, and cost tracking for every LLM generation
SkillCFreshness: C★ 1Claude Code +9 - 0.493score
agents-sdk
zfifteen
Build, run, deploy, and evaluate OpenAI Agents SDK apps from Codex.
SkillCFreshness: C★ 1Claude Code +9 - 0.491score
alpha
zfifteen
Alpha fixture skill.
SkillCFreshness: C★ 1Claude Code +9 - 0.472score
airtable-filters
airtable
Use this skill when the user wants to find, filter, or narrow down Airtable records by field values, even when they don't explicitly say "filter."
SkillCFreshness: C★ 1Claude Code +9
Signals
- Stars
- 1
- Forks
- 1
- Open issues
- 0
- Last commit (this folder)
- Jul 19, 2026 (3 months ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- MIT
- Repository
- github.com/zfifteen/skills-catalog
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from zfifteen
- 0.502score
30-30-30-technical-note
zfifteen
Turn a PGS research finding into a three-layer technical note bundle: (1) Grade-10 plain prose, (2) embedded SVG infographic with concise figures, (3)…
SkillCFreshness: C★ 1Claude Code +9 - 0.498score
agent-bus-collab
zfifteen
Use when the user asks Codex to coordinate with Grok or another local peer agent through Agent Bus MCP, mentions agent-bus, bus topics, readiness handshakes…
SkillCFreshness: C★ 1Claude Code +9 - 0.498score
agy-cli-collab
zfifteen
Sticky Grok→Gemini collaboration via direct local Antigravity CLI (`agy`) subprocess calls only.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
actualize
zfifteen
Reduce reasoning friction and produce the most direct, high-fidelity answer or artifact possible.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
airtable-cli
zfifteen
Lists bases, reads and writes records, manages tables and fields, filters and searches data in Airtable via the `airtable-mcp` CLI.
SkillCFreshness: C★ 1Claude Code +9 - 0.494score
ai-generation-persistence
zfifteen
AI generation persistence patterns — unique IDs, addressable URLs, database storage, and cost tracking for every LLM generation
SkillCFreshness: C★ 1Claude Code +9
Cite this page
agent-vision skill. Agent Mart. https://agentmart.dev/skills/zfifteen-agent-vision (last updated 2026-10-24)