For agents¶
A skill should change the choices in the finished text. Loading one is the beginning of the work.
Use this page to choose a skill, understand its limits and check that the prose carries the intended voice without acquiring new facts.
The skills¶
Install from Fontlab/vexy-fontlab-writing-skills with npx skills add Fontlab/vexy-fontlab-writing-skills. Each skill includes its own rules and references; it requires no sibling skill or external repository. Use current project evidence when the task supplies it.
Choose fontlab-marketing, fontlab-technical or fontlab-neutral for a single register. Use fontlab-write for mixed drafting and fontlab-rewrite for mixed editing. fontlab-tldr condenses a supplied text while preserving its voice and structure through exactly three private rounds, with a target of 20% of the original length.
fontlab-terminology includes a portable glossary snapshot in references/terms.md. The snapshot records naming scope and status, not current availability. fontlab-localization includes rules and examples; live translation tables do not travel with it. fontlab-partners addresses partner content and its commercial constraints.
What they share¶
Every skill carries a house-rules block copied from tools/house-rules.md by tools/scripts/sync_shared.py in the skills repository. Change that source and run the script. Editing a generated copy creates drift at the next sync.
The shared rules govern factual scope, protected text, naming, register, voice and review. The guide's house voice explains the desired effect: precise observation, connected thought, varied pace and room for dry wit where the surface permits it.
How an agent should behave¶
- Establish the audience, task, source facts and output contract. Distinguish a verified fact from a claim inherited from a draft.
- Choose the register per passage and read the relevant skill and its applicable examples.
- Draft or edit one portion at the requested depth. Use a supported detail to make the situation visible. Let the sentence develop the relationship before shortening it.
- Keep literal labels, quotations, code, identifiers and commercial terms intact. Mark unresolved facts precisely in working drafts.
- Run the skill's checklist on that portion, one item at a time, and repair what fails. Repeat for the next portion.
- Review the complete result against the sources, then against the reader's task.
When a passage needs more character, improve the observation first. A sharper noun, a useful aside or a well-placed consequence can do more than added adjectives. An instruction may already have all the character it needs.
When editing, retain an effective source voice. Do not add an invented narrator, personal experience, motive or customer reaction. For TLDR, source voice governs the condensation; do not decorate a plain source to make it resemble another register.
Work in two passes¶
The first pass makes the piece worth reading. Set the opening, develop the paragraphs and choose where the thought lands. Preserve useful irregularities instead of polishing everything to the same cadence.
The second checks the effect of following the text. Would a reader perform the intended action, understand the actual limitation or assess the offer accurately? Remove an analogy's false implication, restore a lost qualification and move a warning before the action.
For a changed skill, try a fresh source packet as well as its bundled examples. Include a plain warning or limitation where expressive rewriting would be harmful. Compare each output claim with the packet and assess the voice by the choices actually made. Repeating a skill's example demonstrates recall, not transfer to another task.
Keep these checks private when the output contract asks for text only. A review report is not an extra deliverable to attach by habit.
Where a tool cannot install skills, the copyable prompts carry the same rules and the same checklist. A separate review prompt gives a fresh chat what it needs to act as the second reader.
Why every skill ends with a checklist¶
A skill that only lists rules relies on the agent to remember all of them while it writes. The evidence says that does not hold for long rule sets. In the IFScale benchmark, the best of twenty models followed 68% of instructions when given 500 at once, and models favored the instructions that came first (Jaroslawicz et al., 2025). The house rules are fewer than that, but each skill adds its own on top, and a draft has to satisfy all of them at once.
Checking works better than remembering, under three conditions.
Each question is answered on its own. Chain-of-Verification has the model plan verification questions about its draft and answer them independently, so one answer does not bias the next; this reduced hallucinations in list and long-form tasks (Dhuliawala et al., 2023). A checklist item such as "every number appears in the evidence" is that kind of question.
The check has something to check against. Iterative self-feedback improved outputs on seven tasks by about 20% on average (Madaan et al., 2023). But a model asked to correct its reasoning without external feedback often fails, and sometimes makes the answer worse (Huang et al., 2024). So the checklist points at the evidence and the text: a passage, a source line, a label. "Does this sound right?" is not a check.
The loop stops. Each skill limits the loop to three rounds, and an item that fails again after its repair is marked rather than attempted a third time. Corrections that pile up in one context leave the failed attempts in view, and a reviewer told to find problems will find some even in a sound draft. Whatever still fails is marked and reported.
Anthropic's guidance for skill authors describes the same pattern for style guides: draft, review against a checklist, note each issue with its location, revise and review again (Skill authoring best practices). Its Claude Code guidance adds that a fresh session reviews more fairly than the one that wrote the draft, and that emphasis such as "IMPORTANT" works on a single line, not on many (Best practices for Claude Code). The skills follow both points: where the agent can open a separate session or subagent, it gives the reviewer the draft, the evidence and the checklist, without its own reasoning. The rules carry their reasons instead of capital letters.
These findings come from benchmarks and engineering guidance, not from measurements of FontLab prose. They justify the method; editorial trials still decide whether a changed skill writes well.
How a person overrides them¶
A direct user instruction takes precedence over a supplied voice sample, which takes precedence over a house default. Adapt punctuation, tone and structure accordingly. Break a default when it would make the writing worse.
A request for a more dramatic style does not establish a more dramatic fact. Keep inventions visibly fictional and unsupported claims unresolved until evidence is available.
Checks¶
In the skills repository:
The first checks shared-rule synchronization, paths and unit tests. The second reports approximate counts and review signals. Stored ranges are comparison data; below 40 words the tool omits range comparisons.
Read a flagged passage in context. Repeated fragments might be an effective acceleration or a string of disconnected claims. A count cannot make that judgment or determine authorship.
Here, ./build.sh check checks publication boundaries, terminology, existing fixture evidence and unit tests. A full build also checks generated links and assets. Add direct editorial trials for changed instructions; passing a build does not show that a skill produces the desired prose.
Adding a skill¶
Create fontlab-<name>/SKILL.md in the skills repository, put its task-specific instructions above the shared-block markers and run the sync script. Include worked examples in references/moves.md.
Give each example a source packet, a weak version, a revision and an explanation of the decision. Include a case where restraint is the right choice. The agent needs to learn when a technique helps and when to leave it out.
End the skill with a checklist of eight to twelve items, each a yes-or-no question about something visible in the text or the evidence, tagged with the house rule it enforces. Choose the failures that skill causes most often; the full rule set is already in the shared block. State the reason for a rule beside it. Keep SKILL.md under 500 lines and move long examples into references/.