"Write it more concisely" is a rule. It has a value in your head and no value at all in the model's. The examples pattern fixes exactly that: stop describing the standard, paste both endpoints of it instead. This issue takes apart four artifacts you can open and counts them — including a countable answer to how many example pairs one rule deserves.
One disclosure first. The roadmap named the "example sections" inside the official brand-guidelines and doc-coauthoring skills as this issue's template. I pulled both and read them line by line: brand-guidelines is 73 lines and contains zero examples; doc-coauthoring is 375 lines and its examples are written for the human, teaching you how to reply — not model-facing sample prose. So the template isn't in those two. It's in the "Examples pattern" section of the official best-practices doc, and in two community skills you can open. And that gap is this issue's first criterion: not every skill should have examples. Every line count and pair count below was counted today from the main branch, not copied from an aggregator.
Copy this one section and you can start: the "Examples pattern" in the official Skill authoring best practices is a complete template. The subject is commit messages, there are exactly three pairs, the form is Input: one line of natural-language description and Output: a code block, and it closes with a single line: "Follow this style: type(scope): brief description, then detailed explanation." The section's own conclusion reads:
Examples convey the desired style and level of detail to Claude
more clearly than descriptions alone.(Per platform.claude.com, that page, 2026-09.) What it kills is a requirement like "write a good commit message" that nobody can accept or reject. Paste three comparable artifacts and the acceptance criterion is sitting right there.
The official skill-creator (anthropics/skills · skill-creator) carries a condensed version of the same section you can copy structurally. It adds one warning: "if 'Input' and 'Output' are in the examples you might want to deviate a little" — don't leave those two literals in the shipped skill, or the model learns the labels along with the style.
blader/humanizer welds them into a four-slot unit, and that unit is the single most copyable thing in this issue.blader/humanizer (MIT, SKILL.md declares v3.0.0) installs with npx skills add blader/humanizer --global; drop --global for the current project only, and it answers to /humanizer. It lists 25 numbered sentence patterns, and every one uses the same four slots: Watch for: (how to spot it) · Problem: (why it counts as a fault) · Before: · After:. The first two slots are the rule, the last two are the example. The rule names the thing; the example draws its border. Pattern 1, Not X but Y, ships this:
Before: It's not just about the beat riding under the vocals;
it's part of the aggression and atmosphere.
After: The heavy beat adds to the aggressive tone.(Per that repo's main-branch SKILL.md, 2026-09.) What it kills is the unfalsifiable judgment "this reads like AI" — with four slots on the page, judgment drops to pattern matching. When not to install it: not as an always-on skill in a technical-docs repo, where the adverbs it strips are the precision you need (Skill 17 covered this).
boraoztunc/skills · stop-slop (393 lines) takes the other route: the pattern lists come first, and a standalone ## Examples block sits at the end with five pairs — each followed by a Changes: line, e.g. "Removed opener, binary contrast structure, and emphasis crutch."
Changes: line, it learns "what got cut was the throat-clearing opener and the binary contrast." It is also the line most people skip when they write their own examples. The move is to copy that three-line unit: Before / After / Changes. When not to install it: it competes with humanizer for the same trigger surface — run both and your voice changes every pass. Pick one.The "Use examples effectively" section of the official Prompting best practices gives three: relevant (mirror your actual use case) · diverse (cover edge cases and, in its own words, "vary enough that Claude doesn't pick up unintended patterns") · structured (wrap them in <example> tags so the model can tell examples from instructions). The recommended count on that page is 3–5 (per platform.claude.com, that page, 2026-09).
anthropics/skills · brand-guidelines is 73 lines with zero examples. Its constraints are #d97757, #141413, Poppins for headings, Lora for body text — a hex code needs no example, because it already is the acceptance criterion. If a constraint's value can be written down in one line, write the rule. Only what resists that gets an example. That's line one of the checklist in section 03.
I counted all 27 second-level sections in blader/humanizer (per main-branch SKILL.md, 2026-09). Across its 25 sentence patterns: 14 carry one pair, 9 carry two, 2 carry three, plus 2 procedural sections (Voice / What to return) with none. Not one pattern exceeds three pairs. Read alongside the official "3–5 examples" guidance, the picture resolves: 3–5 is the total a model can hold for one judgment; per individual rule it lands at 1–3. Patterns that are hard to spot earn two or three; the obvious ones stop at one.
blader/humanizer's README changelog, the v2.9.1 entry records removing "a repeated long example." More pairs is not better; the extra pair is just paying token rent.skill-creator writes this trap into its own body (per main-branch SKILL.md, 2026-09): "if the skill you and the user are codeveloping works only for those examples, it's useless." That is exactly the overfitting you get from polishing two or three examples in a loop — you and the model perfect those passages, and the skill dies on the fourth. The fix is sourcing: take them from drafts you actually sent back and actually revised, then run them past the three official criteria (relevant / diverse / structured).
The official ceiling is 500 lines of SKILL.md body, and examples are the bulkiest, least compressible part of it. The remedy is progressive disclosure: move examples into sub-files and leave a router in SKILL.md. anthropics/skills · internal-comms is the extreme artifact of this: its SKILL.md is 32 lines, and the body is a routing table — by communication type, go read examples/3p-updates.md (46 lines) · examples/company-newsletter.md (65) · examples/faq-answers.md (29) · examples/general-comms.md (15). What it kills is re-explaining your status-report format every single week. How to use it: copy the whole thing into your skills directory and swap those four format files for yours. When not to install it: its whole gathering step assumes Slack / Drive / Calendar MCP access — without those, all you get is formatting.
examples/, but what it holds is not input/output pairs — it's a format spec per communication type. The ## Example Formats section inside company-newsletter.md literally reads "- Announcement 1 / - Announcement 2 / - Announcement 3". It gives you a skeleton, not sample prose. Why? Real internal company comms can't go into a public repository. That's a constraint, not laziness — and it's yours too: don't stuff client or internal drafts into a public repo as examples.anthropics/skills · doc-coauthoring (375 lines) puts all its examples on the input side. Its Step 3, Curation, simply lists four sample replies: "Keep 1,4,7,9" · "Remove 3 (duplicates 1)" · "Remove 6 (audience already knows this)" · "Combine 11 and 12." The text says to request brief justifications in order to learn priorities for the next sections — it uses examples to train the human, so the human's feedback arrives in a form it can learn from. How to use it: any skill that needs repeated feedback from you deserves three sample replies. When not to install it: it's a full co-authoring workflow from a blank page; if you only want a paragraph polished, this is a sledgehammer.
Bill one: overfitting. Anthropic wrote the warning into skill-creator itself (quoted above). Fixes exist: keep examples diverse, and test the finished skill on a task it has never seen.
Bill two: tokens. Examples take up the most room in a SKILL.md, and they're usually what pushes you past 500 lines. The fix is the internal-comms 32-line routing table: examples go to sub-files, and only the relevant one gets read.
blader/humanizer consolidating 35 patterns into 25, then deleting a long example, is that countermeasure in action — an example library expires like anything else, it just doesn't tell you when.brand-guidelines: 73 lines, zero examples, nothing but hex codes and font names. That is the correct call.## Example Formats block in internal-comms is this pattern, not the examples one.Copy these three (specific files, not methodology): ① Structure — the "Examples pattern" section of the official best-practices: three pairs, Input/Output, one "follow this style" line to close. ② Unit — the four slots from blader/humanizer: Watch for / Problem / Before / After, at 1–3 pairs per rule. ③ Annotation — the Changes: line from boraoztunc/skills's stop-slop, naming what each pair actually removed.
How to lay out the directory: follow internal-comms — SKILL.md holds only the routing table, examples live in sub-files. Just don't copy its naming: calling a folder of format specs examples/ has misled plenty of readers.
Explicitly do NOT write examples when: the constraint is a hex code, a font name or a fixed command that fits in one line (brand-guidelines is right to ship zero) · all you have is two drafts by the same person on the same day (diversity fails, overfitting guaranteed) · the examples contain client or internal material and the repo is public (this is exactly why internal-comms ships no real prose).
Three questions before you ship: can a machine verify this constraint? Then write a script. Can its value be written down in one line? Then write a rule. Neither, but you can produce two comparable drafts? Then write an example — and add that Changes: line.
internal-comms is the extreme case.skill-creator writes this warning into its own body.<example> tags, and the 3–5 count (2026-09)## Examples block holds five pairs, each with a Changes: annotation. That line is the thing worth stealingexamples/ sub-files (46 / 65 / 29 / 15 lines); the folder says examples but holds format specs, which is this issue's honest caveat