Type design with agents

Connect OMP using the MCP guide, then ask it to call design_context. This returns the official documentation entry points and a review workflow. Ask it to read this page and your font project’s DESIGN.md. The documentation supplies technique; your brief supplies style. A model name alone does not establish that a drawing is good.

Record the font’s grid, vertical metrics, stroke construction, terminals, contrast, overshoot, spacing goals, and master compatibility requirements. List approved references and explain mark colors. Green, orange, and red have no universal design meaning. For a project that uses green for approved references, orange for revision, and red for replacement, state that explicitly. The agent should preserve green glyphs and their marks unless asked to edit them.

Use glyph_inventory separately on each master to resolve mark labels and Unicode scalars to actual glyph names. U+0038 is decimal 56; do not assume its name is eight. Several glyphs can map to a codepoint. An empty glyph exists but has no outlines or components; an absent glyph has no record at all. Current drawing proposals require an existing glyph record. Create absent records in the editor before requesting outlines.

Good type design is controlled comparison, not point-by-point similarity. An agent should treat the project brief and approved glyphs as evidence, then make one claim at a time and test it in context.

Inspect	Decide	Check
Baseline, x-height, cap height, overshoot, and grid	Which alignments are structural and which rounds must overshoot	Compare straight and round glyphs at the same size
Stems, joins, contrast, and terminals	How stroke logic changes between vertical, horizontal, diagonal, and curved strokes	Compare H, O, n, o, and the target at text and display sizes
Counters, apertures, and dense joins	Whether interior space is still open enough for the intended size and weight	Look at black and white together; equal black area is not equal visual weight
Widths, sidebearings, and recurring pairs	Whether a glyph belongs in words, not merely as an isolated drawing	Proof it in representative words before changing kerning globally
Master structure and component use	Which visual correction can preserve interpolation and which needs an explicit redesign	Run structure checks and compare every affected master

Optical correction is normal: rounds overshoot, joins often need relief, and mathematically equal distances can look unequal. It is not permission to make arbitrary exceptions. Reuse a demonstrated correction from comparable glyphs, state the reason, and leave unusual decisions for human review.

A useful first prompt is:

Read the project design brief and Runebender’s type-design guide. Confirm the Regular and Bold master indices. Use green Regular glyphs and Bold H, O, n, o as references. Inspect their geometry and proof images, then propose Bold a, b, c, d. Preserve the green references. Leave the drawings as proposals and report any interpolation incompatibilities.

The agent should first check whether the target records exist and whether those references are drawn. Compare stem thickness, round-stroke behavior, terminal shapes, counters, and spacing. A single offset or scale factor is not sufficient evidence that the weight fits the family. References from both masters show how the design changes with weight.

propose_edits accepts set_outline: complete contours in UFO point order, with font-unit coordinates and y upwards. A cubic endpoint has type curve and follows two offcurve controls. Closed contours wrap; open contours begin with move. clear_components: true explicitly removes components when replacing a composite with outlines. Width, anchors, encoding, and marks are preserved unless a separate supported operation changes them.

For interpolating masters, retain compatible contour order, starting points, point types, and component structure. The existing structure check is a foreground/proposal check, not a complete family compatibility audit. Compare the result with the other masters before accepting it.

For curve work, ask the agent to inspect U+0038’s direct contours and preserve intentional corners. read_glyph includes tangent angles and signed endpoint curvatures at connected joins. A smooth flag expresses intent; it does not prove continuity. Tangent alignment and matching curvature are separate checks. Degenerate handles need special attention. The reported diagnostics exclude components and identify nonempty contours, not necessarily original point indices. Use the returned point arrays for edits.

The agent can change controls with set_point and smooth intent with set_smooth, then measure the proposed layer again. Inspect the whole curve as well as its joins: a numerically smooth join can still produce an uneven or flat-looking outline. Do not round control points to a preferred grid if that damages the intended curve.

For weight corrections, compare target and references at the same scale. Inspect straight stems, round strokes, counters, joins, and dense regions separately. Do not equate equal black area with equal visual weight. Ask for small batches, such as one orange glyph, before applying the same reasoning to a whole marked group. For red glyphs, explicitly request complete outline replacement and review the resulting structure and spacing.

1. Read current targets and references, including their revisions.
2. Request foreground proofs with explicit glyph names and master indices.
3. Submit a uniquely named proposal with a concrete reason.
4. Proof its layer, then compare with the foreground and reference glyphs.
5. Refine by discarding the draft and proposing from current foreground revisions.
6. Ask for installation only after reviewing the result; re-proof afterward.

MCP proofs contain PNG image blocks and metrics. The client must deliver those images to the chosen model; ask it to describe a visible detail as an initial check. A text-only client cannot perform visual review just because it received an SVG string. Glyph proof sheets show isolated glyphs. The specimen tool adds a bounded basic Latin text proof using harfrust and UFO kerning. See Experiment versions for A/B comparison and PDF export.

proposal_install requires an explicit keep_structure choice. true skips changed point structures. false permits redrawing existing records and can break interpolation. Installation checks the foreground revision again and records one undo step per glyph. A batch can be partly skipped: inspect the returned installed and skipped lists. Installation does not save the font.

Attach the PDF to OMP using its file input, for example @proof.pdf, and ask the agent to use its PDF-reading or rendering tools. Verify that it can see the pages before asking for visual corrections. Runebender’s MCP server does not itself parse PDFs.

Ask for a finding list with page, text or pair, master, observed problem, and suggested correction. Separate outline weight from sidebearings and kerning. Review affected glyphs in other words and pairs before changing spacing globally. Confirm that the PDF corresponds to the current font; a stale proof can describe problems already fixed in the live document.

Outline and advance corrections can use current glyph proposals. Independent kerning versions and basic Latin MCP text proofs are available in the experiment MVP. Automatic creation of absent glyph records and broader text-proof layouts are not implemented yet. Do not describe this version as an automatic full-proof repair system. Keep unresolved findings visible and regenerate the PDF after accepted changes.