Visual Verify

Announce: "I'm using visual-verify to set up a render-vision-fix loop."

NO VISUAL TASK IS COMPLETE WITHOUT RENDERING, SCORING, AND MEETING THE THRESHOLD.

Source code correctness does NOT imply visual correctness. You MUST render to PNG, score with vision (0-10) — Gemini CLI, or a subagent with Read if Gemini is unavailable — and iterate until score >= 9.5. Claiming "done" with a score below threshold delivers broken visuals to the user.

Skipping the score check is NOT HELPFUL — the user gets a visual artifact with defects you didn't verify.

Gemini CLI Vision

Gemini CLI is inherently agentic — it can read files, crop, zoom, and re-examine regions autonomously. No special flags needed for complex vs. simple images; it handles both automatically.

The Loop

0. PAGE MAP -> If Typst + Touying: Skill("teaching:find-slide-page")
       |      Returns heading → physical page mapping
       |      Skip if: single-page file, non-Typst, or page already known
       |
1. CHANGE  -> Modify source code (Task agent)
       |
2. RENDER  -> Produce PNG + PDF (see references/render-commands.md)
       |      Render fails? -> fix source, back to step 1
       |
2.5 TEXT   -> pdftotext pre-screen (Typst/PDF outputs with diagrams ONLY)
       |      Run: pdftotext -f <page> -l <page> -layout output.pdf -
       |      Check for defect #10 (adjacency violations)
       |      Any violations? -> fix source, back to step 1 (skip vision)
       |      Clean? -> proceed to step 3
       |      Skip if: non-PDF output, no diagrams on page
       |
3. VISION  -> Two-part check:
       |      a) INTENT — Does the render match the design intent?
       |         (Does the visual structure argue what it should?)
       |      b) DEFECTS — Scan for the 10 visual defect categories:
       |         1. Text clipped by or overflowing its container
       |         2. Text or shapes overlapping other elements
       |         3. Arrows crossing through elements instead of routing around
       |         4. Arrows landing on wrong element or pointing into empty space
       |         5. Labels floating ambiguously (not anchored to what they describe)
       |         6. Labels squeezed between adjacent elements without clearance
       |         7. Uneven spacing (cramped sections next to spacious ones)
       |         8. Text too small to read at rendered size
       |         9. Parallel sub-diagrams with inconsistent layout
       |        10. Edge labels fused with node text (pdftotext adjacency)
       |      Gemini CLI vision call with SCORING:
       |      → Score 0-10 against checklist items
       |      → Record in SCORES.md
       |
4. DECIDE  -> Score >= 9.5 AND exit criteria met? → DONE
              Score < 9.5 OR defects remain?      → extract fixes, back to step 1

Step 2.5: pdftotext Pre-Screen (Diagrams Only)

What it catches: Edge labels colliding with node text in Fletcher/CeTZ diagrams. When two text elements overlap or touch horizontally in the rendered PDF, pdftotext -layout fuses them into a single string without whitespace — e.g., Republic of1937 war bonds instead of Republic of 1937 war bonds.

When to run: Any page that contains a #fletcher-diagram or #cetz.canvas. Skip for pages with only body text, tables, or bullet points.

How to check:

1. Extract the page: pdftotext -f <page> -l <page> -layout output.pdf -
2. Collect all node labels and edge labels from the source code
3. Scan the extracted text for any two labels appearing concatenated without whitespace
4. A fused pair (node text running directly into edge label text) is a BLOCKING defect — return to step 1 without spending a vision call

Why this works: PDF text extraction preserves spatial positions. If elements have clear space between them, whitespace appears in the output. If they overlap or touch, words concatenate. This is cheap, deterministic, and catches the most common Fletcher layout failure that vision scoring misses.

Why this is NOT a replacement for vision: pdftotext only catches horizontal text collisions. It cannot detect labels sitting on top of arrow lines (vertical overlap), misaligned arrows, or spacing imbalance. Vision (step 3) is still required for a full pass.

When to Stop

The loop is done when ALL of these hold:

• Score >= 9.5
• The rendered output matches the design intent (not just defect-free but correct)
• No text is clipped, overlapping, or unreadable
• All arrows/lines connect to the right elements and route cleanly
• Spacing is consistent and composition is balanced
• You'd show it to someone without caveats

Don't stop after one clean pass just because there are no critical bugs — if the composition could be better, improve it. Conversely, don't loop forever on cosmetics — 5 iterations max before escalating to the user.

Invocation

Set a /goal whose condition gates on the enumerable defect substrate — fused labels and BLOCKING visual defects — not the 0-10 aesthetic score (an LLM vision score is noisy and won't stably hit 9.5; chasing it is a treadmill). Run the loop body (render → vision → fix) inside each turn. The evaluator reads the defect findings from the transcript.

/goal Visual Task N [TASK NAME] is complete when the rendered output at
[OUTPUT PATH] has ZERO BLOCKING defects (clipping, overlap, illegible text,
wrong connections) and the pdftotext pre-screen finds no fused labels
(diagrams only), per SCORES.md. The look-at 0-10 score is advisory, not the
gate. Stop after 5 turns.

Getting it set — you cannot Skill(goal) (it is a built-in UI command), and printing the line is a no-op (slash commands dispatch only on the user input path). Only the top-level session may run:

bash ${CLAUDE_SKILL_DIR}/../craft/scripts/goal-self-send.sh '/goal <condition>'

Proceed only if the helper exits 0 (goal-self-send: delivered via …) or the user explicitly confirms activation. Otherwise hand the literal condition to the user and STOP. A spawned agent returns that literal command to its caller without running the helper. Once active, each subsequent turn fires automatically until zero BLOCKING defects remain (not until an aesthetic score crosses a bar). With no active goal there is no loop at all — one pass, no gate. After the terminal PASS, clear it with bash ${CLAUDE_SKILL_DIR}/../craft/scripts/goal-self-send.sh '/goal clear'. Proceed only after that command exits 0 or the user explicitly confirms the goal is cleared; otherwise print /goal clear and stop.

Score Tracking

Initialize SCORES.md before the first iteration:

# Visual Verify Scores

| Iteration | Score | Threshold | BLOCKING | COSMETIC | Delta |
|-----------|-------|-----------|----------|----------|-------|

Each vision call must score the output 0-10:

• 10.0 = all checklist items pass, zero issues
• 9.5 = 95% pass, 1-2 cosmetic issues remain (default threshold)
• < 9.0 = BLOCKING issues present

The score reflects the fraction of checklist items that pass. Gemini counts BLOCKING and COSMETIC issues against the domain-specific checklist, and the score = (items passing / total items) * 10.

Vision Calls

Use look-at's look_at.sh wrapper for all vision calls.

# Single backend (Gemini CLI, default)
"${CLAUDE_SKILL_DIR}/../../skills/look-at/scripts/look_at.sh" \
    --file "/tmp/visual-verify.png" \
    --goal "[CONTEXT-ENRICHED GOAL]"

# Consensus mode (Gemini + GPT-5.4 in parallel)
"${CLAUDE_SKILL_DIR}/../../skills/look-at/scripts/look_at.sh" \
    --file "/tmp/visual-verify.png" \
    --goal "[CONTEXT-ENRICHED GOAL]" \
    --consensus

Use --consensus for diagram pages. Use single backend for non-diagram pages.

If Gemini CLI fails (API key unavailable, missing, or API error), DO NOT fall back to pdftotext AS A REPLACEMENT for vision. pdftotext cannot detect vertical overlap, misalignment, or spatial defects — it only extracts text. Using it as the sole visual check produces false passes.

Instead, spawn a subagent that uses the Read tool directly on the rendered PNG:

Agent(
    prompt="Read the image at /tmp/visual-verify.png and score it. [CONTEXT-ENRICHED GOAL]. Score 0-10 against the checklist. Report BLOCKING and COSMETIC issues separately.",
    description="Vision fallback: score rendered PNG",
    subagent_type="general-purpose",
    # Blocking: verification cannot resolve without the score. A backgrounded agent returns a
    # completion notification, not a verdict.
    run_in_background=false
)

Why this is safe: The Iron Law against Read on images exists to protect the main conversation's context window (images cost 1000-5000 tokens). Subagent context is throwaway — the tokens are discarded when the agent returns its short summary. The cost rationale does not apply.

pdftotext's role: pdftotext is prohibited as a replacement for vision (the lecture 22 incident: it missed CeTZ vertical text overlap). But it is required as a pre-screen in step 2.5 for diagram pages — it catches horizontal adjacency violations that vision frequently misses. The two prongs complement each other: pdftotext catches fused text; vision catches spatial/vertical issues.

Multi-Prong Verification

For pages containing diagrams (Fletcher, CeTZ), verification uses two complementary prongs that run in sequence. Neither prong alone is sufficient.

Prong 1: pdftotext adjacency check (step 2.5) — cheap, deterministic

Catches: edge labels fused with node text, labels running into each other horizontally. Misses: labels sitting on top of arrows, vertical overlap, spatial imbalance.

# 1. Extract the page
pdftotext -f <page> -l <page> -layout output.pdf /tmp/vv-text.txt

# 2. Collect expected labels from source
#    Grep all node(..., [...]) and edge(..., label: [...]) content from the .typ file

# 3. Check for fused pairs
#    Scan extracted text for any two labels concatenated without whitespace
#    e.g., "Republic of1937" = node "Republic of China" fused with edge "1937 war bonds"
#    Any match = BLOCKING, return to step 1 without spending a vision call

How to detect fused text: For each edge label in the source, check whether it appears in the pdftotext output concatenated with an adjacent node label — no whitespace between the end of one and the start of another. The specific pattern is: <end of node text><start of edge label> or <end of edge label><start of node text> appearing as a single unbroken string.

Prong 2: Gemini vision (step 3) — expensive, catches spatial issues

Catches: labels on top of arrows, vertical overlap, arrow misrouting, spacing imbalance. Misses: subtle horizontal adjacency (Gemini often scores these 9/10 or higher).

Prong 3: Multi-model vision consensus (step 3b) — second opinions

For diagram pages, use look-at's --consensus mode (runs Gemini + GPT-5.4 in parallel) plus a Claude subagent. Run all three in parallel:

# Gemini + GPT-5.4 consensus (via look_at.sh)
"${CLAUDE_SKILL_DIR}/../../skills/look-at/scripts/look_at.sh" \
    --file "/tmp/visual-verify.png" \
    --goal "[CONTEXT-ENRICHED GOAL]" \
    --consensus

# Claude subagent (reads PNG natively; inherits the session model)
Agent(
    prompt="Read the image at /tmp/visual-verify.png. [CONTEXT-ENRICHED GOAL]. Score 0-10. Report BLOCKING and COSMETIC issues.",
    description="Vision prong: Claude",
    subagent_type="general-purpose",
    # Blocking: the consensus rule needs this score alongside the look_at.sh output.
    run_in_background=false
)

When to use prong 3: Always for diagram pages. For non-diagram pages (body text, tables), prong 2 alone is sufficient.

Calibration note (measured with Claude Opus; re-baseline if scores drift): Gemini tends to underscore overlap defects — marking BLOCKING issues as COSMETIC and scoring 8/10 when Claude scores 5/10 on the same image. Trust the stricter score. If any model flags BLOCKING, treat it as BLOCKING regardless of other models' severity ratings.

Consensus rule: If any model flags a BLOCKING issue that others missed, treat it as a real defect and fix it. A 9.5 from Gemini + a 7.0 from Claude = not clean. All models must agree the output is clean.

The prongs feed into each other:

• If prong 1 finds fused text → fix immediately, skip prongs 2-3 (save the vision calls)
• If prong 1 is clean → run prong 2 for spatial verification
• If prong 2 scores >= 9.5 but low confidence → run prong 3 for consensus
• If any prong finds issues → fix, then re-run ALL prongs on next iteration
• All prongs must pass for the page to be scored as clean

Goal Assembly

NEVER call Gemini with a generic goal. Goals must reference the spec, checklist items, and prior feedback.

For diagrams (fletcher, CeTZ): use describe-first strategy, NOT judgment prompts.

• Step 1: Ask Gemini to transcribe all text elements with [FULL | PARTIAL] annotations
• Step 2: Claude diffs transcription against expected elements from source code
• Step 3 (optional): Run spatial/overlap check for non-clipping issues
• Why: Judgment prompts ("is X visible?") invite yes-man answers. Transcription forces the model to report what it actually sees — clipping becomes objectively detectable via text mismatches.

See references/goal-templates.md for full copy-paste templates per domain.

Translating Non-Python Feedback

Before editing Typst source to apply fixes, load the tinymist skill: Skill(skill="tinymist:typst") It has Fletcher/CeTZ reference docs with correct parameter names (label-side, label-pos, spacing, inset). Without it, you will guess at parameter names and waste iterations.

Gemini returns structured pixel measurements. Claude translates to source code:

Complex Diagrams (3+ Failed Iterations)

If the same spatial issue persists after 3 iterations, escalate to the reference sketch approach: have Gemini draw an ideal layout in matplotlib and translate coordinates.

See references/complex-diagram-strategy.md for the full approach.

Quick Render Reference

See references/render-commands.md for the full reference.

Red Flags

When NOT to Use

• One-off visual checks: Use Gemini CLI directly, not the full loop
• Text-only verification: Use standard dev-accept
• Compilation checks only: Just run the compile command

Reference Files

• references/goal-templates.md -- copy-paste goal templates per domain
• references/render-commands.md -- render commands for all supported domains
• references/rationalization-prevention.md -- verification facts (with drive consequences) and red flags
• references/complex-diagram-strategy.md -- reference sketch approach for persistent layout failures
• references/examples.md -- worked examples (Typst slide, matplotlib chart, diagram escalation)