AI-assisted font proofreading
AI-assisted font proofreading is a guided review of the font that is open in Glyphs. The agent gathers evidence, helps prioritize it, opens useful review views, and proposes the next check. Glyphs remains the source of truth and the designer remains responsible for design decisions.
The goal is not to generate a specimen or a report document. The useful output is a reviewed font, a traceable set of findings, and confidence that approved changes were checked again.
Glyphs MCP already provides many of the focused tools used in this process, including font inspection, visual rendering, spacing review, kerning review, OpenType inspection, outline analysis, annotations, and guarded apply tools.
A resumable end-to-end proofreading session and the remaining family-wide analyzers are planned. See the implementation plan.
What the AI contributes
An AI agent is useful when it acts as a review coordinator rather than an automatic type designer. It can:
- establish the exact font, master, instance, selection, and intended coverage
- run repeatable checks and distinguish failures from unusual-but-valid design
- reduce large result sets to the cases most likely to deserve attention
- create targeted Glyphs tabs and visual comparisons
- relate symptoms that may share one cause, such as spacing and kerning findings
- explain evidence in typographic language
- propose bounded changes through dedicated tools
- remember accepted risks and rerun checks affected by an approved change
AI cannot infer the intended character set, language support, or design intent reliably from outlines alone. It also cannot decide whether an unusual form is a mistake. Those inputs must come from the designer.
Evidence model
Every finding should identify its target and the kind of evidence behind it.
| Evidence | Meaning | Appropriate conclusion |
|---|---|---|
| Deterministic | A reproducible rule or computation failed. | Report an objective problem with exact targets. |
| Statistical | A value differs materially from related glyphs or styles. | Request review; do not assume the outlier is wrong. |
| Visual | The rendered result appears uneven, ambiguous, or inconsistent. | Record a design observation for human judgment. |
Useful severities are blocker, error, warning, and note. Severity does
not replace the evidence type: a visual concern remains a design note unless a
separate reproducible failure supports it.
Before the review
Tell the agent:
- which open font to review
- whether the target is a static font, family, or variable font
- intended languages or a required glyph-set preset
- required styles, instances, and OpenType features
- whether a previous release should be used as a regression reference
- whether the session is inspection-only or may propose confirmed changes
If intended coverage is unknown, the agent can review existing glyphs but should mark character-set completeness as not assessed.
Recommended sequence
1. Establish context
The agent should begin with read-only context tools such as list_open_fonts,
get_selected_font_and_master, get_font_masters, and get_font_instances.
It should state the chosen font and master before interpreting results.
A future proofreading session will also record a source fingerprint. If the font changes, findings based on the older state become stale rather than being silently treated as current.
2. Review family structure
Inspect masters, axes, instances, naming, vertical metrics, alignment settings, export flags, and custom parameters. The agent should separate contradictory or invalid configuration from optional production recommendations.
Questions for the designer include:
- Do axis extremes and defaults match the intended family?
- Are all required instances present and exporting?
- Are family and style names suitable for the intended distribution channel?
- Are metrics and parameters intentionally shared or different between styles?
3. Review coverage and glyph construction
Check glyph names, Unicode assignments, export state, required layers, components, anchors, and outline structure. Review related glyphs together so that construction differences have context.
Use get_glyph_details, path and component inspection tools, Unicode review,
and focused Glyphs links. Large glyph sets should be split into meaningful
groups rather than returned as one unbounded response.
4. Review interpolation
For multi-master and variable fonts, check compatibility before judging visual progression. Mechanical compatibility includes contour topology, node order, components, and anchors. Visual review then samples named instances, axis extremes, defaults, and selected intermediate locations.
An incompatible outline is an objective finding. A weight or width progression that feels uneven is a visual finding for the designer.
5. Review spacing
Run spacing review before kerning. Start with representative control glyphs, then investigate exceptions across uppercase, lowercase, figures, punctuation, symbols, diacritics, and composites.
Use review_spacing for measurements and proposals. Use
the native curve and candidate Reporters or Glyphs Edit tabs for visual rhythm. If a change is
useful, run apply_spacing as a dry run and wait for explicit approval before
using confirm=true.
6. Review kerning
Check group membership and group-level behavior before proposing glyph-glyph
exceptions. Use generate_kerning_tab to create focused worklists and
review_kerning_bumper to find collisions and near-misses.
Every kerning finding should name the effective pair, master, group or exception source, measured gap, and a reproducible test string. See the kerning workflow.
7. Review OpenType behavior
Inspect feature code and the glyphs involved with show_opentype_features,
list_style_sets, and related feature tools. A complete workflow also shapes
representative text with relevant feature, language, direction, and variation
settings.
The review should cover features actually promised by the project. It should not treat the absence of an optional feature as an error.
8. Conduct focused visual passes
Use short, purposeful review rounds instead of one undifferentiated specimen:
- lowercase texture and word rhythm
- uppercase and mixed-case spacing
- paragraphs at intended text sizes
- tight and loose tracking
- figures, currency, fractions, and punctuation
- diacritics and composed/decomposed sequences
- similar or easily confused characters
- style and weight progression
- axis extremes and intermediate variable-font locations
- recently changed glyphs
- the full exporting glyph set in manageable groups
Use the native curve and candidate Reporters. The removed PNG renderer is useful when the AI client needs pixels with metrics, sidebearings, bounds, nodes, handles, anchors, or guides. See visual glyph review.
9. Inspect temporary binaries
Source checks cannot prove that an exported font behaves correctly. A complete review therefore exports representative builds into an isolated temporary directory and inspects the actual binaries.
The planned companion process can use:
- Fontspector for general OpenType QA
- fontTools for targeted table and variation inspection
- HarfBuzz for deterministic shaping
- Diffenator3 for optional comparisons with a previous build
Temporary fonts should never be installed as system fonts, and binary tools must not modify the Glyphs source.
10. Triage, change, and verify
Group findings by probable cause. For each proposed correction, the agent should show the evidence, exact targets, expected effect, and dry-run result.
After approval:
- Use the narrowest dedicated mutation tool.
- Read back the changed state.
- Rerun checks affected by the change.
- Reopen the relevant visual comparison.
- Mark the finding verified only when its original evidence no longer reproduces.
The agent must not save the font unless the designer explicitly requests it.
When the review is complete
A review is complete when every required stage is recorded as passed, resolved, accepted risk, or skipped with a reason. There is no opaque overall score.
Completion does not mean every visual note was changed. It means the designer has seen the relevant evidence and made an explicit decision.
Prompt template
Use Glyphs MCP to conduct an evidence-led font review of the font currently open in Glyphs.
Start read-only. Confirm the font, masters, instances, intended Latin language coverage, and required OpenType features.
Review in this order:
1. family structure
2. glyph coverage and construction
3. interpolation, if applicable
4. spacing
5. kerning
6. OpenType behavior
7. focused visual comparisons
8. exported-binary QA, if the companion tools are available
For every finding, distinguish deterministic, statistical, and visual evidence. Open the relevant glyph or review tab in Glyphs. Propose the smallest safe correction, but do not mutate until I approve a dry run. Verify approved changes and never save automatically.