Unconventional Coding Skills

Twenty-eight themed coding skills that change how an AI writes code, each one a philosophy, a set of reusable patterns, and a checkable spec.

The Skills

The current release contains exactly 28 independently installable skills — 13 short-form poetry skills plus 15 added classical, Asian-adjacent, and Western forms:

Skill	Identity	Standalone activation
choka	Long alternating verse with a closing couplet	explicit choka or long-metered-verse request
dodoitsu	Four-line 7-7-7-5 folk form	explicit dodoitsu or 7-7-7-5 request
gogyohka	Free-form five-line code	explicit gogyohka or free-form five-line request
haibun	Narrative code with a three-line poetic landing	explicit haibun or narrated prose-and-verse request
haiku	Compact three-line 5-7-5 code	explicit code-haiku or 5-7-5 request
katauta	Directly addressed three-line 5-7-7 fragment	explicit katauta or 5-7-7 fragment request
lunes	Three-line 5-3-5 form with a short middle	explicit lune or 5-3-5 request
monoku	Complete program on one physical line	explicit monoku or one-line-program request
renga	Linked stanzas with visible handoff pivots	explicit renga or chained-stanza request
sedoka	Two mirrored 5-7-7 stanzas	explicit sedoka or two-stanza-response request
senryu	Human-nature humor in three 5-7-5 lines	explicit senryu or human-comedy request
sijo	Three long lines ending in a twist	explicit sijo or three-line-twist request
tanka	Five-line 5-7-5-7-7 result and reflection	explicit tanka or 5-7-5-7-7 request
kyoka	Comic tanka: five lines 5-7-5-7-7 ending in a joke	explicit kyoka or comic-verse request
somonka	Paired exchange of two 5-7-5-7-7 verses	explicit somonka or paired-exchange request
bussokusekika	Six-line 5-7-5-7-7-7 verse with a sealing verdict	explicit bussokusekika or 5-7-5-7-7-7 request
imayo	Four lines with a rolling 7-5 pulse per line	explicit imayo or 7-5-7-5 song request
kanshi	Four lines of paired 7-7 couplets with a turn and resolve	explicit kanshi or couplet-verse request
zappai	Three-line moment freed from haiku's kigo rules	explicit zappai or free three-line moment request
waka	Classical five-line 5-7-5-7-7 scene, turn, and resolve	explicit waka or classical verse request
renshi	Chain of linked short stages passing a torch	explicit renshi or linked-relay request
sonnet	Fourteen lines in quatrains with a volta and couplet	explicit sonnet or 14-line verse request
villanelle	Nineteen lines in tercets with two repeating refrains	explicit villanelle or 19-line refrain request
cinquain	Five lines shaped 2-4-6-8-2, a pyramid with a two-token landing	explicit cinquain or 2-4-6-8-2 request
ryuka	Four-line Okinawan 8-8-8-6 song with a short landing	explicit ryuka or 8-8-8-6 request
fibonacci	Lines growing 1-1-2-3-5-8, each the sum of the previous two	explicit fibonacci poem or golden-ratio request
limerick	Five lines ~8-8-5-5-8 in AABBA rhythm with a comic punchline	explicit limerick or AABBA request
etheree	Ten lines climbing 1-2-3-4-5-6-7-8-9-10 to the result	explicit etheree or ten-line ladder request

Every skill is self-contained and carries its own activation contract, minimum requirements, cross-language examples, and helper policy. Installing one skill does not require downloading or loading any other skill or repository file.

Installation

Packaged .skill files

cd skills
python3 package_skills.py            # writes dist/<name>.skill for all 28
python3 package_skills.py --target .agents/skills   # or install directly

Copy the directories

cp -r skills/terry-davis ~/.agents/skills/   # repeat for each skill

• Codex: .codex/skills/
• Gemini CLI: .gemini/skills/
• Freebuff: place in .agents/skills/

Loading / Usage

skill({ name: "terry-davis" })
skill({ name: "psych" })
skill({ name: "haiku" })
// ... any of the 28 current skills

COMPREHENSIVE_DOCUMENTATION.md is a historical four-skill reference; use each current SKILL.md as the source of truth for activation and examples.

Evaluation

• static_skill_audit.py, objective SKILL.md quality audit (frontmatter, scope, requirements, language coverage, helper policy, and no-mock-code integrity). The latest validation run scores 1.00 on all 28 skills (see CURRENT_SCOPE_RELEASE_REPORT.md); earlier 0.79-0.80 results are historical snapshots.
• Latest evaluation (2026-08-19) — see CURRENT_SCOPE_RELEASE_REPORT.md: post-hardening model arms lifted same-model (Mistral) shape convergence 15 → 20/28 and strict one-shot passes 0 → 3/28; the agentic loop reached 3/28 strict and 19/28 shape (ceiling confirmed at 6 generations); the independent blind-routing score is 244/250 (0.976); a hands-on skill-test-kit pass solved 13/13 tasks with correct output and form-checker passes; current CI is 37/37 and the current historical suite 7/7.
• The external evals-infra/ repository contains the with-skill vs. baseline harness (scaffold → grade → aggregate → viewer). Installed skills do not require the harness at runtime. See its HOW_TO_RUN_EVALS.md when validating a source checkout.

⚠️ The historical 180-query trigger set and its recorded decisions predate the standalone activation rewrite. They remain reproducible regression data, not independent evidence that the revised descriptions generalize. The external harness is the source of truth for future independently authored routing evaluations.

GitHub Actions CI (per-skill repos)

Each skill can live in its own repository (mrfentmen/<skill>). Every push to any of them runs the Audit & Package workflow (ci/audit-and-package.yml, distributed by push_all_skills.sh into .github/workflows/):

1. Static audit (static_skill_audit.py), scores SKILL.md on 9 dimensions. Below AUDIT_MIN_SCORE (default 0.75) the run fails red and nothing ships, bad quality doesn't publish.
2. Repackage (package_skills.py), rebuilds dist/<skill>.skill from the current tree (CI machinery excluded).
3. Artifacts, SKILL_AUDIT.json + the .skill upload on every run, even failures, for diagnostics.
4. Auto-commit, on the default branch, refreshed artifacts are committed back with [skip actions] so the loop can't re-trigger.

The audit + packager scripts are layout-aware: they run identically in the monorepo (skills/) and inside a per-skill repo (SKILL.md at the root).

Deploy: existing repos were seeded before CI existed, so re-run GITHUB_TOKEN=<pat> bash push_all_skills.sh once to install the workflow

• scripts into all 28 skill repos (the "name already exists" message is expected and harmless; it still pushes the new .github/ files).

The infra repo (mrfentmen/skills-infra)

The evals-infra/ harness lives in its own repository so the standalone skill repos stay lean. GITHUB_TOKEN=<pat> bash push_infra_repo.sh creates or updates that infrastructure repository. Generated reports, caches, workspaces, and build artifacts remain outside the standalone skill payloads.

Safety

Unconventional ≠ broken. All skills must still produce working code, contain no malware or exploit material, and keep theatrics in the style layer.

License

Provided as-is for educational and creative purposes. Use responsibly.