Cultural Relic Drawing

Convert artifact photos, descriptions, or existing drafts into archaeological drawings that prioritize scientific record over visual drama.

Default to the highest-fidelity output the evidence supports. If the source is incomplete, lower the ambition of the drawing instead of inventing hidden structure, exact thickness, missing ornament, or a fake scale bar.

Source Conditioning / Calibration Gate

Classify the source before planning the drawing. This gate decides the maximum defensible output tier.

• Measured object + calibrated multi-view
  • Source includes real measurements or trustworthy calibration plus multiple reliable views.
  • Eligible for report-grade candidate.
  • Final report-grade measured drawing still requires human review, proportion checking, and publication preparation.
• Near-orthographic view with scale reference
  • Source includes a scale or clearly trustworthy size reference and a near-orthographic face or profile.
  • Eligible for a single-face orthographic drawing and possibly report-grade candidate of that face only.
  • Do not extend report-grade claims to unseen sides, thickness, or unsupported sections.
• Multi-view but uncalibrated
  • Source includes several views, but no reliable calibration or measurement basis.
  • Eligible only for evidence-bounded AI technical draft.
  • Never present this directly as report-grade measured drawing.
• Single oblique photo
  • Source is one angled or partial photograph without trustworthy calibration.
  • Eligible only for visible-face redraw within evidence-bounded AI technical draft.
  • Never reconstruct a full orthographic object record from this alone.
• Text-only description
  • Source is verbal description with no reliable image evidence.
  • Eligible only for view plan, prompt package, or schematic concept.
  • Never present the result as a normative archaeological drawing.

If more than one class seems plausible, choose the weaker class unless calibration and geometry support the stronger claim unambiguously.

Output Tiers

State or internally determine the output tier before prompt-writing.

• Tier 1: report-grade measured drawing
  • Reserved for measured or calibrated evidence plus human review.
  • AI may assist with drafting, cleanup, or layout, but this is not a one-pass pure generation output.
• Tier 1.5: report-grade candidate
  • Allowed when evidence is strong but not yet fully measured, calibrated, or publication-finalized.
  • Must state what is still missing, such as calibration, contour verification, line cleanup, or publication preparation.
• Tier 2: evidence-bounded AI technical draft
  • Default tier for most photo-based requests.
  • Intended for draft production, review, correction, vectorization, and later publication preparation.
• Tier 3: archaeological-style illustration
  • Allowed when evidence is weak, schematic, or primarily communicative.
  • Must not be presented as a report drawing.

Workflow Modes

These modes sit underneath the evidence gate and output tier. They control execution only; they never upgrade the allowed output claim.

• main-structure pass
  • Responsible for overall contour, major structure, projection, baseline symmetry or asymmetry, and broad ornament zoning.
  • Must intentionally suppress or defer high-risk local detail such as scripts, dense repeated ornament, glare-heavy edges, openwork crossings, and damaged boundaries.
• high-risk local pass
  • Responsible only for isolated high-risk regions, such as scripts or symbols, dense ornament clusters, openwork intersections, reflective edges, break lines, and other ambiguity hotspots.
  • Must not alter the global contour or re-decide the overall projection.
  • Must not imply a separate exported local image unless the user explicitly asks for one.
• review-and-merge pass
  • Responsible for reconciling the main structure draft with local corrections, demoting unsupported detail, and issuing the final Risk note.
  • Must decide whether weak local detail is merged, simplified, omitted, or explicitly downgraded.

Relationship rules:

• The Source Conditioning / Calibration Gate decides the maximum defensible claim.
• The Output Tier decides how authoritative the deliverable can be.
• The Workflow Mode decides how execution proceeds.
• Workflow modes can make draft production safer and more auditable, but they cannot strengthen evidence or upgrade the output tier.

Preprocessing / Region Triage

Run this triage before main-structure pass. It is not optional.

• Choose one primary geometry reference.
  • Use it to lock contour, projection, and the main face or profile.
• Choose any secondary detail references.
  • Use them only to confirm local detail that is compatible with the same observed structure.
• Crop or mentally isolate the relevant object, face, or local area before prompting when the source includes excess background, glare zones, or unrelated annotation.
• Name the high-risk regions before generation.
  • Typical categories: scripts or symbols, dense repeated ornament, openwork crossings, reflective edges, damaged or missing boundaries, severe overlap or occlusion.
• Decide whether the task should run as:
  • whole-object or whole-face drawing only
  • high-risk local handling only
  • split workflow with both
• Demote or ignore low-value inputs instead of averaging them into the result.
  • Typical demotion cases: blur, heavy glare, bad perspective, strong occlusion, shallow depth of field, or poor local focus.

Minimum triage outputs:

• source class
• output tier
• primary geometry reference
• secondary detail references
• named high-risk regions
• selected mode sequence

If the triage is weak or contradictory, lower the ambition of the drawing before writing prompts.

Mode I/O Contract

Treat each workflow mode as a contract with clear inputs, outputs, and prohibitions.

• main-structure pass
  • Input:
    • source class
    • output tier
    • primary geometry reference
    • object or face scope
  • Output:
    • stable global contour
    • projection and baseline
    • major structure and broad ornament zoning
    • unresolved local-risk list
  • Must defer:
    • scripts or symbols
    • dense ornament micro-topology
    • micro-breaks and tiny damaged edges
  • Must not change:
    • evidence gate
    • output tier
• high-risk local pass
  • Input:
    • named high-risk regions
    • locked main-structure draft
    • compatible secondary detail references
  • Output:
    • per-region local decisions and localized corrections only
  • Must not change:
    • global contour
    • global orientation
    • overall projection
    • output tier
• review-and-merge pass
  • Input:
    • main draft
    • local-pass decisions
  • Output:
    • merged draft
    • downgrade decisions
    • final Risk note
  • Must report each named local region as:
    • resolved
    • simplified
    • omitted
    • left uncertain

User overrides such as 只跑主结构, 只做高风险局部, or 只做合并复核 may restrict the mode sequence, but they must not bypass the evidence gate, output tier, or Risk note rules.

Workflow

1. Establish the evidentiary ceiling before drawing.

   • Prefer image editing with the source artifact photo as a reference when a photo exists.
   • Assign roles to the references before prompting: choose one primary geometry reference for the overall contour and view, and treat other images as secondary detail references for local relief, ornament, wear, or inscriptions only.
   • Prefer tightly cropped, well-focused images of a single artifact. Demote or ignore blurry, low-contrast, strongly reflective, or perspective-heavy photos instead of averaging them into the result.
   • Decide the Source Conditioning / Calibration Gate class first, then cap the allowed Output Tier before writing prompts.
   • Distinguish report-grade measured drawing, report-grade candidate, evidence-bounded AI technical draft, and archaeological-style illustration.
   • If the source is a single oblique or partial photo, limit the output to the visible face or visible profile rather than fabricating a full orthographic reconstruction.
   • Treat inscriptions, seals, stamped marks, maker's marks, numerals, alphabetic characters, monograms, ligatures, punctuation, and any text-like or symbol-like traces as high-risk structure. Never let the model translate, modernize, regularize, autocomplete, normalize, or typeset them.
   • Distinguish markings on the artifact from text or symbols outside the artifact. Preserve on-object scripts, letters, numbers, and symbols only when they are clearly visible and diagnostically important; remove museum labels, captions, watermarks, inventory numbers, and background writing unless the user explicitly asks to document them.
   • Never invent exact dimensions, wall thickness, or section geometry from an unrelated view.
   • Complete Preprocessing / Region Triage before entering main-structure pass.
   • Detect whether the artifact contains high-risk regions such as scripts or symbols, dense repeated ornament, openwork crossings, strong specular glare, damaged or missing boundaries, or severe overlap or occlusion.
   • If no high-risk region is present, skip high-risk local pass.
   • If high-risk regions are present, name them explicitly and route them into high-risk local pass.
   • Default to automatic mode switching, but honor an explicit user override such as 只跑主结构, 只做高风险局部, or 只做合并复核.
   • Use this evidence matrix:
     • Single oblique or partial photo: allow only a visible-face archaeological drawing or a cautious technical illustration.
     • Single clear side/front photo with near-orthographic view: allow one visible orthographic view of that face only.
     • Multiple clear photos of opposite sides: allow a two-side plate with matched scale and aligned baselines.
     • Multiple views plus genuine evidence for lip, break, section edge, or thickness: allow section or profile information only where that evidence exists.
     • Real dimensions from the user or trusted metadata: allow a truthful scale bar; otherwise omit it.

2. Choose the view system that matches the artifact.

   • For rotationally symmetric vessels, use the standard left-half section and right-half exterior view only when profile and thickness are actually supported by the source.
   • For asymmetrical or complex artifacts, use multiple aligned orthographic views.
   • For openwork plaques, fittings, pendants, and similar flat artifacts, use the frontal orthographic face as the main view and add side or section information only when thickness is evidenced.
   • Keep all views in orthographic alignment. Do not allow perspective, camera angle, or dramatic foreshortening.

3. Encode the drawing language as hard constraints.

   • Use black lines, dots, hatching, and white negative space only.
   • Maintain a strict hierarchy: overall contour first, major structural boundaries second, diagnostic ornament third, incidental surface noise last or omitted.
   • Use a fixed light logic from upper left at roughly 45 degrees.
   • Use thin lines on lit convex edges and heavier lines on shadow-side convex edges.
   • Reverse that rule for incised or recessed features.
   • If script, letters, numbers, punctuation, or symbol strokes are present on the object, treat them as observed shapes, not readable text. Copy their visible geometry exactly when legible; if illegible, keep them as uncertain marks rather than replacing them with normalized modern text or symbols.
   • Use dashed lines only for hidden, missing, or reconstructed features when there is an evidentiary basis for inference.
   • Do not use painterly shading, wash effects, photographic shadows, gradients, or decorative background.

4. Write prompts in the correct order.

   • State the non-negotiable scientific constraints first.
   • State the chosen output tier near the beginning so the model is not free to over-claim precision.
   • State the current workflow mode explicitly near the beginning.
   • State the artifact type, view system, and orientation second.
   • State the line, point, and shading logic third.
   • State layout, delivery shape, scale-bar handling, risk-note requirement, and any required labels last.
   • Inject output conventions into the prompt itself: white background unless the user requests another plate ground, centered artifact, clean margins, legible line hierarchy at publication scale, and restrained labels.
   • Specify solid lines for visible structure, dashed lines only for hidden or reconstructed structure, and section fill or hatching only when a real section is justified.
   • If the user asks for a plate or provides multiple views or artifacts, specify a grid or aligned plate layout with consistent baselines and truthful scale handling.
   • Ban perspective, color, studio lighting, glossy reflections, speculative restoration, and decorative embellishment explicitly.

5. Review the result against the source before accepting it.

   • Check the overall contour and silhouette first. If the outer shape drifts, reject early.
   • Check the major structure next: view choice, voids, openwork, appendages, sections, and aligned baselines.
   • Check observed damage, wear, asymmetry, and missing areas next. If the result looks cleaner or more complete than the source, revise toward the source.
   • Check dense ornament and repeated motifs at zoomed-in local crops rather than only at full-plate scale. Look for breaks, merged loops, duplicated motifs, and invented closures.
   • If high-risk local pass was used, record whether each named local region was resolved, simplified, omitted, or left uncertain.
   • Check line logic last: hierarchy, convex-versus-concave treatment, dashed-line meaning, and whether the drawing still reads as technical rather than decorative.
   • Then review the image against the acceptance checklist in references/standards.md.

Non-Negotiable Rules

• Treat absolute fidelity to the original object as the highest rule.
• Prefer orthographic projection over pictorial realism.
• Record what is visible; infer only when the inference is standard, minimal, and supported.
• Separate observed form from reconstructed form.
• Preserve breakage, asymmetry, wear, and irregularity when they are visible in the source.
• Add a scale bar only when real dimensions are known. If dimensions are unknown, omit the scale bar or reserve space for later annotation, but never fabricate numeric scale.
• Never let a stronger output claim exceed the evidence gate.

View Selection

Choose the narrowest view set that the evidence supports:

• Symmetric vessels: use the conventional half-section system only when profile and thickness are genuinely supported.
• Asymmetric vessels: orient the main view to the most diagnostic side and add projected views only when needed.
• Stone tools: use front, back, and side views when morphology requires them.
• Openwork or flat ornaments: prioritize the visible face and preserve silhouette plus voids exactly.
• Sculptural objects: prefer left/right orthographic side views before expanding to more views.

Read references/standards.md for the full view-and-layout rules, artifact-specific conventions, and the acceptance checklist.

Prompting Guidance

When using an image model:

• Prefer reference-image editing over text-only generation.
• Tell the model to convert, redraw, or regularize into archaeological drawing, not to reimagine or stylize.
• Tell the model exactly which visible structures must be preserved.
• Tell the model what it must remove from the source photo, such as background cloth, labels, museum captions, glare, or cast shadow.
• When inscriptions or symbol-like markings matter, explicitly say treat all characters, letters, numerals, and symbols as image geometry, not as text content and do not translate, simplify, regularize, replace with fonts, or repair missing strokes.
• Explicitly say scientific record first, artistic expression second.

When the first pass is weak:

• Tighten geometry constraints before adding more stylistic language.
• Re-state the reference hierarchy explicitly: follow the primary reference for contour and projection; use secondary references only to confirm local detail already visible on the same face or view.
• Ask for less shading, fewer invented details, and stricter orthographic alignment.
• Reduce the scope from full reconstruction to visible-face archaeological drawing when evidence is limited.
• For dense ornament, ask the model to separate outer contour, main motif, and ground pattern with different line importance rather than rendering all details at equal strength.
• If the result still over-claims certainty, demote the output tier rather than piling on more stylistic language.

Read references/prompt-templates.md when you need ready-made prompts for vessels, stone tools, openwork ornaments, or review passes.

Read the matching example only when a task fits it closely:

• examples/vessel-profile.md for vessel profiles or half-section requests.
• examples/stone-tool.md for stone tool multi-view drawings.
• examples/openwork-plaque.md for openwork plaques, fittings, pendants, and flat ornaments.

Execution

After the prompt is ready, choose an execution path that matches the output tier.

• Prefer the strongest image-generation model the current environment can call, especially one that supports reference-image editing and iterative revision.
• Prefer image editing from the source photos over text-only generation.
• Do not hardcode a local provider or skill. Choose the most capable available model at runtime.
• If multiple callable models exist, prefer the one that best preserves structure, contour, and damage while allowing prompt-based correction.
• Default to automatic mode switching:
  • Start with main-structure pass.
  • Invoke high-risk local pass only when a named high-risk region exists.
  • Finish with review-and-merge pass.
• If the user explicitly requests a subset mode, honor it, but keep the existing evidence gate, output tier, and Risk note requirements in force.
• Make mode transitions explicit in user-facing progress updates.
  • At start, state the full triage record, not just the mode name.
  • The opening update should include:
    • source class
    • output tier
    • primary geometry reference
    • named high-risk regions
    • selected mode sequence
    • current mode as main-structure pass
  • If local risk exists, state high-risk local pass and name the affected regions explicitly.
  • At the end, state review-and-merge pass and provide the final Risk note.
• Tier 2: evidence-bounded AI technical draft
  • AI generation may proceed directly.
  • Generate at least one correction pass when the first result violates the evidence ceiling, orthographic logic, or line-drawing conventions.
• Tier 3: archaeological-style illustration
  • AI generation may proceed directly, but the prompt and deliverable must not imply report-grade authority.
• Tier 1 and Tier 1.5
  • AI may assist with drafting, but the output must go through review-and-refinement steps after generation: contour check, proportion check, view correction, line cleanup, and publication preparation.
  • Model availability alone never justifies report-grade measured drawing or report-grade candidate.
• If recurring work follows one house style, the highest-leverage improvement is not prompt length but domain adaptation: a small paired dataset of photos and accepted line drawings can support LoRA-style customization better than generic prompting alone.
• After generation, lightly clean the result if needed: remove isolated speckles and mend tiny discontinuities, but do not simplify or fuse fine diagnostic strokes.
• If no image-generation model is callable, return a prompt package, view plan, and review checklist rather than pretending the drawing was produced.

For Codex/OpenAI environments, UI metadata may exist in agents/openai.yaml. Claude Code does not need that file to trigger the skill; it relies on SKILL.md and the frontmatter description.

Deliverables

Draft output

• Default to a single PNG on a white background unless the user explicitly asks for another format or plate style.
• If the user provides multiple views of one artifact, default to one composed plate with aligned views rather than unrelated separate outputs.
• If the user provides multiple artifacts or explicitly asks for a plate, compose as a grid with consistent margins, aligned baselines where appropriate, and truthful scale cues only when measurements are known.
• If the user requests separate exports, keep the main composed plate plus individual views only when that split clearly helps publication or review.
• Draft output is suitable for review, markup, local correction, and later cleanup.

Publication output

• If the user’s goal is publication, report production, catalogue work, or figure submission, prefer vectorizable or production-ready deliverables when the workflow actually supports them, such as SVG, PDF, high-resolution TIFF, or layered line-art source.
• Treat AI output as an initial line-art draft unless it has also gone through human review, correction, and publication preparation.
• Do not promise automatic vectorization or layered source output unless the current toolchain can genuinely produce it.
• Keep labels, numbering, and scale bars restrained and publication-oriented. Omit numeric scale when dimensions are unknown.

Risk note

• Every output must end with a short Risk note.
• For strong evidence, keep the note brief and state the residual limits.
• For weaker evidence, name the uncertain areas explicitly, such as worn ornament, low-contrast relief, reflective glare, dense motif topology, damaged edges, or ambiguous scripts/symbols.
• If high-risk local pass was used, say whether each named local area was resolved, simplified, omitted, or left uncertain.

LoRA Readiness

This section is a future-facing extension. It is not required for normal use.

Current backend reality:

• This skill does not currently provide a built-in trainable image-generation backend.
• If the available image backend is only a hosted general model, treat LoRA as external future work rather than something the skill can execute directly.
• In OpenAI-only image-generation environments, assume workflow control, reference discipline, and review passes are the primary optimization tools unless a separate trainable image backend is introduced.

Consider LoRA only when all of the following are becoming true:

• the workload is recurring rather than occasional
• a stable house style or review standard exists
• the same classes of failure keep recurring
• a small paired dataset is available

Minimum recommended data structure:

• input artifact photo
• accepted line drawing
• same view only
• no mismatched pairs such as oblique source photo plus idealized full reconstruction

Useful optional metadata:

• artifact type
• view
• source class
• output tier
• risk regions

Suggested collection buckets:

• base set
• ornament hard set
• script/symbol hard set

Do not train on examples such as:

• blurry or reflective photo paired with a clean idealized reconstruction
• ancient or damaged script paired with normalized modern text
• broken ornament paired with fully repaired ornament

Treat this as training-readiness guidance only. Do not assume a full training pipeline, commands, or tooling are already available inside this skill.

Failure Modes

Avoid these common errors:

• Perspective distortion disguised as technical drawing.
• Symmetrizing an asymmetric object.
• Replacing observed damage with idealized edges.
• Converting ancient, foreign, symbolic, or damaged markings into normalized modern characters, letters, numerals, or punctuation.
• Presenting an evidence-bounded AI technical draft as report-grade measured drawing.
• Letting high-risk local pass overwrite the global contour or projection established by main-structure pass.
• Treating a local-pass correction as evidence strong enough to upgrade the output tier.
• Inventing unseen backs, interiors, or thickness.
• Using uniform line weight everywhere.
• Rendering the image like fantasy concept art, engraving art, or decorative poster art.
• Adding a fake scale bar or fake measurement marks.