Anti-pattern handles for Claude Code. Each Aesop fable is one Claude Code skill that names a failure mode the model already recognizes from training. Dropping the handle into a prompt — or letting the skill auto-trigger on description match — activates a precise "don't do X / watch for X" schema that English has a worn-smooth phrase for but no operational checklist.
Sibling catalog to chengyu-skills. Where chengyu mostly name strategic disciplines ("do Y"), Aesop fables mostly name anti-patterns ("don't do X"). The two are complementary, and the real leverage is in cross-catalog composition — e.g. chengyu-retreat-is-best authorizes a retreat, aesop-sour-grapes guards the post-mortem against rationalizing it.
git clone https://github.com/evnchn-agentic/aesop-skills.git ~/aesop-skills
~/aesop-skills/install.shinstall.sh symlinks each skill directory into ~/.claude/skills/ (symlinks, not copies, so repo updates are reflected live). Restart Claude Code or start a new session for discovery. Uninstall: ~/aesop-skills/install.sh uninstall removes only the symlinks this repo created.
Three high-confidence entries — each names a failure mode with no existing chengyu coverage:
aesop-cry-wolf(Perry 210) — false alarms have asymmetric credibility cost; gate warnings on real confidence or the channel stops being readaesop-sour-grapes(Perry 15) — don't let a failure re-rate the goal as unworthy; it corrupts the post-mortemaesop-golden-goose(Perry 87) — don't destroy a renewable source (rate limits, operator attention, maintainability) for one task's throughput
More candidates are filed as issues rather than shipped — the catalog ships few, distinct entries and lets the debatable ones cook. This is a deliberate brick (抛磚引玉): a small, honest MVP that invites refinement.
An Aesop fable earns a skill only if it passes the overlap test: does it name a failure mode with no existing coverage in the chengyu catalog (or this one)? That test — not "is it a cautionary tale" — is what does the real work. Several iconic fables were cut for redundancy, not weakness:
- The Tortoise and the Hare → collapses into
鐵杵磨針(compound effort over a long horizon) - The Crow and the Pitcher → sits between
鐵杵磨針and化整為零(accumulate / decompose)
A charming fable that duplicates an existing handle degrades the catalog. Distinctness over coverage.
This repo is Aesop specifically — fables in the Perry Index, the standard catalog of the Aesopic tradition. Famous lookalikes that are not Aesop are excluded and belong to other (possible future) catalogs:
- The Wolf in Sheep's Clothing — Biblical (Matthew 7:15) and medieval (Abstemius, 15th c.), not Aesop. Its disguised-adversary / prompt-injection schema is genuinely useful, but it isn't ours to claim here.
- The Scorpion and the Frog — 20th-century Russian origin (c. 1933), not in the Aesopic corpus at all.
Bucket by source, not by alphabet.
The handle uses the short English idiom when the phrase stands alone in normal speech (aesop-cry-wolf, aesop-sour-grapes, aesop-golden-goose — people say all three). When the short form isn't a standalone phrase (nobody says "dog reflection"), use the full title (aesop-dog-and-its-reflection). Either way, pack the literal failure-mode keywords ("false alarm", "alert fatigue", "rate limit", "post-mortem") into the description field — that field is the auto-discovery trigger surface.
- Create
aesop-<handle>/SKILL.mdfollowing the shape of the existing entries: fable + Perry Index number, schema, "when this applies", behavior, anti-pattern, relationship to the chengyu catalog, moral. - Apply the overlap test before writing. If a chengyu already names it, don't add it.
- Re-run
install.sh(idempotent). Use it for a week before committing to the handle.
- Tested on Claude Code (macOS, Linux). Skills follow the agentskills.io spec.
- Aesop's fables are public domain in every jurisdiction; quoted/retold freely.
- MIT (see
LICENSE).