Focused Claude Code plugins

The marketplace packages instruction discovery, conditional Codex context, adversarial review, and parser-aware docstring help as narrow plugins. The work is less about adding more prompting than controlling when each piece of context enters the model's working set.

My role

Independent project · plugin design and contract testing

GitHub

System map

Session event

Narrow matcher

Targeted context

Model action

Regression contract

Agent tooling

Principle: Context only when relevant
Scope: One problem per plugin
Contracts: Observed hook payloads

Context is a budget

Persistent instructions are easy to add and hard to reason about. Every unrelated token competes with the repository, user request, and evidence the model needs for the current task.

Each plugin therefore owns one intervention and stays silent outside it: discover instructions only when a file path crosses the project boundary, explain Codex only when that plugin is enabled, and add LangChain-specific rules only when the target actually defines a tool.

Probe the hook system instead of guessing

The repository includes a capture harness for the real payloads Claude Code sends when a user invokes a slash command, a model invokes a skill, or a file is read. That exposed materially different event paths and namespaced plugin command names that a documentation-only implementation would have missed.

Deterministic tests replay crafted payloads through the hooks so matching, deduplication, and graduated context behavior remain visible without spending a model call.

Keep review scope mechanically narrow

The adversarial-review plugin separates a small request builder from the review agent's actual operating contract. The builder captures the exact slice, collection guidance, and user concern; the reviewer owns the skeptical stance, evidence bar, and output shape.

That boundary prevents a request helper from silently redefining review policy and makes it harder for a broad repository scan to leak unrelated findings into a scoped review.

Instruction discovery by content, not path

The discovery plugin can load CLAUDE.md and AGENTS.md files outside the project tree, then deduplicates identical instructions by content hash. Worktrees, clones, and copied templates stay quiet while genuinely different context remains available and changed files trigger a re-read nudge.

Engineering choices

The decisions that shaped it.

The implementation details matter because each one closes a specific failure mode or keeps an important boundary visible.

Make silence a feature

A plugin that is irrelevant to the current task contributes no standing context and no background behavior.

Test observed events

Hook matchers and payload fields are based on captured runtime behavior, then protected by deterministic contract tests.

Separate request from policy

Skills gather a precise target; agents retain ownership of the review or transformation contract applied to it.

Where it landed

A public marketplace of focused Claude Code extensions whose main product constraint is predictable context: the right instruction, at the actual trigger, and nothing when it is irrelevant.

Next case study

dotfiles

Keep reading