Visual Checker

Figure Checker

Visual QA sub-agent for AI-generated images. It compresses the image, reads it with vision, evaluates against scene-specific checklists, and outputs a structured ## Figure QA Report with pass/fail verdict and regeneration guidance.

When To Use

• You just generated a figure and need to verify quality before returning it
• You are running a generation pipeline (e.g., academic-paper, rhetoric-of-decks) and need automatic visual QA
• You want structured, actionable feedback for regeneration

Input

Prepare three values before invoking:

Scene guidance:

• academic — checks flat vector style, clean background, readable labels, publication-quality aesthetics
• slides — checks text size for projection, contrast, layout balance, no overflow
• general — checks content fidelity, visual quality, artifacts, text rendering, proportions

Invocation via herdr

The visual-checker runs as a separate pi sub-agent in a herdr pane. Use four steps:

Step 1 — Split a new pane

{
  "action": "pane_split",
  "direction": "down",
  "newPane": "figure-qa"
}

Step 2 — Run the sub-agent

Use the invoke.sh helper to avoid shell escaping issues with multiline content:

{
  "action": "run",
  "pane": "figure-qa",
  "command": "bash ~/.agents/skills/visual-checker/scripts/invoke.sh '<image_path>' '<scene>' '<intent>'"
}

With optional extra context:

{
  "action": "run",
  "pane": "figure-qa",
  "command": "bash ~/.agents/skills/visual-checker/scripts/invoke.sh '/tmp/fig.png' 'academic' 'Pipeline overview of our method' 'This is Figure 1 in the paper'"
}

What invoke.sh does internally:
Loads the system prompt from scripts/prompt.md, then calls:
pi --print --model axonhub/gemini-3.1-pro-preview --thinking off --tools read,bash --no-skills --no-context-files --no-extensions --no-session --system-prompt "..." "<user input>"

Step 3 — Wait for the sub-agent to finish

{
  "action": "wait_agent",
  "pane": "figure-qa",
  "statuses": ["done", "idle"],
  "timeout": 120000
}

If it times out, read whatever output is available and surface the partial result.

Step 4 — Read the QA report

{
  "action": "read",
  "pane": "figure-qa",
  "source": "recent-unwrapped",
  "lines": 100
}

Locate the ## Figure QA Report section in the output.

Output Format

## Figure QA Report

**File**: <image_path>
**Scene**: <academic | slides | general>
**Verdict**: ✅ PASS / ⚠️ MINOR ISSUES / ❌ REGENERATE

### Summary
<One sentence on overall quality>

### Issues Found

| Severity | Issue | Location |
|----------|-------|----------|
| Critical | <e.g. garbled text in title> | top-left |
| Minor    | <e.g. slightly off color>    | background |

### What's Good
- <genuine positives>

### Regeneration Guidance
<If REGENERATE: specific prompt edits to fix the issues.
 If PASS or MINOR: "No regeneration needed" or "Optional: ...">

Handling Results

Batch QA

For checking multiple images (e.g., all slides in a deck), reuse the same pane:

run → wait_agent → read → run → wait_agent → read → ...

No need to split a new pane between checks.

Pane Cleanup

After all checks are done, stop the pane:

{
  "action": "stop",
  "pane": "figure-qa"
}

Troubleshooting

Files