jonZlotnik

jonZlotnik/how-to-agent

Plugin Claude Code

Plugin marketplace listing 1 plugin: how-to-agent.

jonZlotnik/how-to-agent

Plugin Claude Code

A library of Agent Skills (SKILL.md open standard) — composable lessons on decision-making, debugging, hygiene, defensive design, Unix-style composition, and technical writing. Portable across any skills-compatible agent. Adapted from Ada Jane Anderson's pragmatic-programmer-skills.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Fixes, files, or deliberately accepts each small defect — never ignores one. Use when entering code you will work in, noticing decay (stale TODOs, skipped tests, dead code), or leaving a mess 'for later'. Do not use to justify unscoped rewrites.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Replaces metaphor and decorative wording with literal, specific language. Use when writing or reviewing technical prose — docs, comments, commit messages, error text. Do not use for deliberately narrative writing such as talks or essays.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Stops the program loudly at the first impossible state instead of continuing on corrupt data. Use when writing error handling, defensive checks, retries, or reviewing code that swallows exceptions. Do not use for expected, recoverable errors.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Identifies the real need behind a vague task before designing. Use when given a terse feature request or a 'just do X' instruction, or when the request names a solution rather than a need. Do not use when the need is already explicit.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Finds knowledge encoded in more than one place and consolidates it. Use before adding similar functionality, when one change forces edits in three places, or when reviewing schemas, configs, and docs. Do not use to merge coincidental lookalikes.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Drives debugging from evidence to root cause instead of blaming tools, libraries, or users. Use when investigating any bug, test failure, crash, or unexpected behavior, before proposing a fix. Do not use for deciding who is responsible.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

States the conclusion or instruction first, then the supporting detail. Use when structuring a document, section, comment, commit message, or reply. Do not use where sequence is the content, such as step-by-step tutorials.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Turns a vague 'something feels off' into a named, testable concern before shipping. Use when unease survives 'looks fine to me', or when tests pass and that surprises you. Do not use to delay after the doubt has been checked and disproved.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Deletes words that add no information. Use when revising any prose before publishing or committing it — docs, comments, commit messages, error messages, UI text. Do not use where cutting would remove precision or a necessary qualification.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Splits sentences and paragraphs so each carries a single idea. Use when a sentence chains clauses with 'and', 'which', or dashes, or when a paragraph covers several topics. Do not use where joining two ideas expresses a real dependency between them.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Measures coupling by changing one thing and counting what else must move. Use when designing a module, reviewing a PR that touches many files, or when a small change cascades through unrelated areas. Do not use on disposable prototype code.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Rewrites passive sentences so a named actor performs the action. Use when drafting or reviewing docs, comments, commit messages, or error messages. Do not use when the actor is unknown or irrelevant and naming one would mislead.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Builds a disposable prototype to answer one question, then discards it. Use when exploring an unfamiliar design space, library, algorithm, or UX, or before a hard-to-reverse commitment. Do not use when requirements and approach are already known.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Checks every unverified belief the work depends on instead of assuming it. Use when stuck debugging, before merging a risky change, or when 'it should work' appears in your reasoning. Do not use to re-prove what a test already covers.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Weighs how cheaply a decision can be undone before committing to it. Use before framework, vendor, schema, API, or format choices, or when a decision is treated as final without examination. Do not use for choices that are already cheap to undo.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Designs programs whose output serves as another program's input. Use when designing a tool, CLI, or service interface, or when data is reachable only through an interactive UI or private binary format. Do not use for single-caller throwaway scripts.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Generates repetitive code from the rule that describes it. Use before hand-writing many similar functions, fixtures, or configs, or when a schema already contains the answer. Do not use when the cases genuinely differ.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Matches conventions users already know instead of inventing new ones. Use when designing a CLI, API, or library surface, picking flag names, defaults, or exit codes, or reviewing a 'creative' convention. Do not use where no existing convention fits.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Models a repeatable declarative domain as a small language instead of ad-hoc code. Use when a config file is accumulating operators, conditionals, or templating, or when writing another bespoke parser. Do not use for single-use configuration.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Requires a measured bottleneck before performance work. Use when something feels slow and tuning is tempting, when reviewing a performance PR, or when a design is justified on speculative speed grounds. Do not use to reject profiled optimizations.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Moves complex branching into data so the logic stays simple. Use before writing a long if/elif chain on string or enum values, or when control flow mirrors the rows of a spec table. Do not use when there are only a few stable cases.

jonZlotnik/how-to-agent

Skill Claude CodeCodex

Accepts reasonable input variation and emits strictly well-formed output. Use when designing a parser, deserializer, or API boundary, or reviewing code that trusts upstream input. Do not use to silently repair input that indicates a real bug.