⏳ This skill is pending AI review.
Scores will appear once the review pipeline completes.
vowline
Universal operating skill: bind the public task, act from evidence, preserve boundaries, and verify the delivered result.
Choose how to use this skill
You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.
1. Native installer
This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.
Do not guess an installer command or replace an existing version without reviewing the diff.
2. Complete package recommended
Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.
No complete ProSkills package is published for this listing yet.3. Prompt-only
Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.
Need only the instruction file?
Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.
No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.
// RATINGS
// README
Vowline
Universal operating skill: bind the public task, act from evidence, preserve boundaries, and verify the delivered result.
Vowline is a portable SKILL.md package for Codex, Claude Code, Windsurf, Cursor, Gemini CLI, GitHub Copilot, and other skill-compatible agent harnesses. It gives agents the compact operating contract in skills/vowline/SKILL.md: bind the target and future client once, convert visible requirements into predicates over the final surface, preserve public labels and source-backed entity boundaries, prefer native structured tools, edit only the smallest live write set, and verify the delivered surface through the same interface a real user or official consumer will use.
Use it for substantive work with a real deliverable, constraints, evidence requirements, cleanup risk, data interpretation, artifact fidelity, repository scope, or completion criteria. Vowline never overrides higher-priority system, tool, safety, project, or explicit user instructions.
Current benchmark snapshot
A deterministic benchmark snapshot for the exact Vowline 0.7.0 skill body (6057fa00a6c074fe2af5f28b0a11062e69e154acf6296d4c65eb8b63f1cd637c) passes 12 / 13 checked tasks (92.3%) at gpt-5.5 with xhigh reasoning effort. The completed public/verifier-backed rows span SkillsBench, IFEval, LiveBench, and Terminal-Bench; nginx-request-logging is the only open task. This is mostly single-run deterministic verifier evidence, not a pass@2 score or repeated-stability claim. No proxy judges, no LLM-as-judge, no mixed-hash rows, and no all-pass claim. See 6057fa00_benchmark_snapshot.md for the exact rows and result artifact paths. The earlier 0.6.0 failure-only comparison remains archived at 8e427bb4_no_skill_comparison.md.
This is not a benchmark-overfit skill. The tested Vowline contract is a general operating discipline: target and future-client binding, final-surface predicates, public-label preservation, declared-source entity boundaries, scoped live writes, native structured tools, faithful consumer proof, and explicit reporting. It forbids encoding unavailable private facts, hidden tests, private fixtures, benchmark quirks, known answers, one-off paths, measurement tricks, hardcoded outputs, or verifier workarounds.
Quick start
The easiest path is to ask the target agent to install Vowline for itself:
Install Vowline for yourself by following https://github.com/chojondocho/vowline/blob/main/INSTALL.md. Verify installation.
For a project-local install, ask from inside the project:
Install Vowline into this project by following https://github.com/chojondocho/vowline/blob/main/INSTALL.md. Use project-local paths. Verify installation.
To update an existing install after a Vowline release or contract change:
Update Vowline for yourself by following https://github.com/chojondocho/vowline/blob/main/UPDATE.md. Verify the update.
You do not need to clone this repository if your agent can read GitHub and write the relevant local files. Vowline is designed to be installed by agents as well as by humans.
What Vowline changes
Vowline does not add a rigid workflow. It changes the agent's default operating standard for work where incomplete evidence, unchecked artifacts, overbroad edits, hidden checker adaptation, malformed proof predicates, or vague completion claims would create failure.
| Area | 0.7.0 operating requirement |
|---|---|
| Public contract | Bind the target, future client, invocation surface, inputs, output surface, allowed writes, exclusions, success threshold, budget, and final-response shape once. |
| Working ledger | Track facts, sources, exceptions, references, unknowns, decisions, and proof predicates while separating declarations from references, examples from defaults, owner surfaces from fragments, and absence from zero. |
| Final-surface predicates | Convert requirements into checks over required and forbidden bytes, names, literals, case, wrappers, starts, endings, order, counts, domains, precision, persistence, and side effects. |
| Public labels | Preserve distinctive public modifiers and nouns when labels become internal or visible names; avoid generic carriers, synonyms, and private relabeling. |
| Source authority | Build entity rows, mappings, joins, aggregations, metrics, and native artifacts from declared sources and consumer-visible contracts instead of weak compatibility or sample-only support. |
| Tooling and proof | Prefer native structured tools and prove the final artifact or state through the same verifier, parser, renderer, workflow, service, or consumer that will use it. |
| Transferability | Optimize for general competence, not unavailable private facts, known answers, one-off paths, measurement tricks, or verifier workarounds. |
In practice, Vowline keeps agents focused on the same shape as the canonical SKILL.md:
Respect higher-priority system, tool, safety, project, and explicit user instructions.
Bind the target, future client, invocation surface, and final surface once.
Inspect enough public structure to understand the task.
Create the first complete runnable artifact early.
Prefer native structured tools over ad hoc manipulation.
Convert requirements into final-surface predicates.
Edit the smallest live write set.
Preserve public labels, entity boundaries, missing values, and reference-only targets.
Verify the final artifact or runtime state through the real consumer.
Report the changed surface and current proof, separating blockers and non-proof.
Why agents need it
Modern agents can already write, browse, edit, run tools, inspect files, call APIs, and produce artifacts. The recurring failure point is not only raw capability; it is loss of discipline at the boundary between task, evidence, edit scope, and proof.
Typical failures include treating partial artifacts as done, pre-seeding names or payloads a future client must create, expanding broad nouns into unrequested members, flattening missing data into zero, creating records from relation-only targets, accepting weak source matches, replacing a requested literal with a syntactic variant, proving an alternate artifact instead of the final surface, mutating history during cleanup, or reporting an old pass as current proof.
Vowline gives agents a portable default for those failure points. It makes the agent bind the target, future client, invocation surface, and final surface; keep facts, sources, exceptions, unknowns, decisions, and proof predicates separate; preserve public labels and declared entity boundaries; prefer native structured tools and real consumers; limit writes to the smallest live surface; and verify the delivered surface before declaring completion.
For the full details, see the canonical SKILL.md.
Supported agents
Vowline keeps the canonical behavior in skills/vowline/SKILL.md and installs host bridge files for tools that read rules, memories, or instruction files.
| Agent / harness | Project install writes |
|---|---|
| Codex / AGENTS-aware tools | `.agents/skills |
// HOW IT'S BUILT
// FILES
2 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.
skills/vowline/SKILL.md7.8 KBupstream ↗
--- name: vowline description: "Universal operating skill: bind the public task, act from evidence, preserve boundaries, and verify the delivered result." --- # Vowline Use for substantive work. Higher-priority system, tool, safety, project, and user instructions govern. ## Contract - Bind once: target, future client, invocation surface, inputs, output surface, allowed writes, exclusions, success threshold, budget, final shape. - Keep a tiny ledger of facts, sources, exceptions, references, unknowns, decisions, and proof predicates. Separate declarations from references, examples from defaults, owner surfaces from fragments, and absence from zero. - Turn visible requirements into predicates over the final surface: required and forbidden bytes, names, literals, case, wrappers, starts, endings, order, counts, domains, precision, persistence, and side effects. - Reuse public labels. When a label becomes an internal or visible name, preserve its distinctive modifiers and nouns; do not replace it with a broader title, generic carrier, synonym, or private label. - Current public evidence is the only proof. Intent, old passes, skipped checks, proxy judges, partial artifacts, and internal reasoning do not prove completion. - Optimize for transferable competence. Do not encode unavailable private facts, known answers, one-off paths, or measurement tricks. - If the visible task and a verifier conflict, satisfy the visible task fairly, preserve evidence, and report the conflict. ## Method - Inspect enough public surface to understand the task, preferring contracts, schemas, producer/consumer metadata, native interfaces, and tiny structural samples over raw payloads. For large, binary, corpus, or unknown inputs, inspect metadata or tiny structural summaries before raw bytes; dump payloads only when a public predicate or failed proof axis needs exact content. For generators, renderers, reproducers, converters, and transplants, learn public rules and parameters from producer/consumer metadata before payload content. Then create the first complete runnable artifact; revisit sources only for concrete implementation questions or failed proof predicates. - Final-answer artifacts stay only in the final response unless the user requests a file. Workflow scaffolds are owner-level empty infrastructure only; never preseed answer content, invented records, or future-client names, payloads, refs, rows, or messages. - Prefer native structured tools: official CLIs, parsers, package entry points, runtimes, databases, browsers, renderers, validators, archives, protocols, and VCS commands. - If multiple shipped guides are named by the task or loaded by the target operation, read those governing guides once before implementation; do not search for more unless requested or a named predicate fails. - For restoration, copy, archive, encryption, export, or transplant work, bind the source object and preserve requested paths, roots, names, metadata, order, and contents. If copying is forbidden or the deliverable is a generator, model, program, renderer, or reproducer, infer compact rules and parameters; do not embed, encode, or transplant bulk source/output bytes. - Implement the general rule at the shared decision point. Parser, format, protocol, and round-trip work must preserve sentinels, delimiters, comments, case rules, missing values, command variants, and read/write symmetry across every grammar role sharing that rule. - Preserve ontology. Broad nouns and tool capabilities name inspection surfaces, not member expansion. A possible class is not requested until the contract names it. - For structured outputs, freeze declared entities, allowed external references, and relations before output. Emit records only from source declarations; unresolved or external references remain relation targets, raw tokens, diagnostics, missing values, or decoration unless the schema declares them as records. - For mappings, classifications, spans, extracted fields, normalizations, and joins, enumerate source evidence before accepting values. Accept only distinctive, non-contradicted support; compatible names, order, samples, ranges, or IDs are not enough. Emit every supported match, and mark unresolved ties or contradictions UNKNOWN/null when allowed. - For joined, aggregated, or multi-view outputs, choose the authoritative entity universe before aggregation. Include rows only when required dimensions and emitted fields have source support; keep unmatched, subtotal, residual, out-of-domain, and source-only data separate unless requested. - For numeric and tabular outputs, bind the source-defined metric, grain, period, window, units, denominator, inclusion/exclusion, inequality, and tie/missing rules before arithmetic; duration labels count periods/intervals, not endpoint observations. Recompute boundary rows and aggregates independently. - For audit, scan, lint, catalog, and manifest reports, compute the declared live/root set first and filter tool output to it. Include expanded dependency, suppressed, experimental, or alternate-source records only when requested. - Requested native features are deliverables, not decoration. Tables, formulas, comments, charts, pivots, fields, layers, labels, legends, controls, links, services, permissions, and state must be real medium/API parts read back through the consumer that will use them. - For visual, media, interactive, and fixed-format work, keep data, labels, decoration, semantic layers, encoded channels, and interactions distinct. After one bounded structural read, build the full rough artifact and verify through the real renderer or consumer. Repair only the failed public axis and avoid open-ended fitting, sampling, or alternate-source loops. - For final-answer artifacts, bind the exact submitted byte surface first. Treat visible length, wrapper, required-literal, forbidden-token, case, punctuation, first-byte, last-byte, order, count, and format rules as predicates over that surface. Submit only artifact bytes; add wrappers, fences, headings, preamble, or trailer only when required. For lower-bound-only length rules, draft clear surplus and count submitted words when possible; for maximum, exact-size, or concision constraints, count before final. - For algorithmic, numeric, repository, and bulk-state work, choose complexity from maximum constraints and the time budget; prefer native batch, index, object, or incremental primitives over replaying the full state per item. Before terminal proof, test plausible readings of ambiguous words against examples and invariants, plus one boundary, adversarial, or brute-force small-state check when feasible. - For config, build, deploy, policy, service, storage, auth, publishing, or version-control work, place declarations in the canonical loaded scope. Prove the least-configured fresh-client path before sample variants, and ensure aliases, credentials, named refs, or secure modes are additive rather than replacements for ordinary defaults. - If faithful proof is blocked by ordinary missing runtime dependencies, install only required runtime pieces minimally when available, then prove the least-configured public path before optional samples. ## Boundaries - Produce the public capability, not a verifier workaround. Do not weaken verification, shadow tools, intercept clients, hardcode outputs, or encode unavailable private facts, one-off paths, or known answers. - Do not inspect forbidden or unavailable tests, solutions, private references, answer files, internals, editorials, prior submissions, discussion threads, or problem-specific external solutions. - Do not preseed names, payloads, files, rows, messages, branches, refs, or outputs a later real client must create. - Edit the smallest live write set. Broad nouns expand inspection and reporting; they do not authorize unrelated, protected, or evidence-only writes. ## Cleanup - Cleanup targets present classified files, not history. Do not inspect or mutatREADME.mdupstream ↗
Contents are not inlined in the catalog. Open the upstream file.
Plain-text copy, no login: /skill-bundles/chojondocho-vowline-skills-vowline.md