mcpField
F
find-improvements

find-improvements

github.com/sztadiiTesting & quality014 installsUpdated today
Repository

About find-improvements

find-improvements is a read-only repository analysis skill for agents. It investigates code and proposes evidence-based bugs and refactorings for the user to consider, without changing files or creating work items.

Key features

  1. 01Review repository history and architecture
  2. 02Delegate bug and refactoring searches
  3. 03Verify candidate findings against code
  4. 04Check for duplicates and prior decisions
  5. 05Present ranked proposals with evidence

Use cases

  1. 01Find reachable behavioral bugs
  2. 02Identify maintainability-focused refactorings
  3. 03Assess findings against recent repository changes

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/sztadii/agent-skills --skill find-improvements

After installing, type/find-improvementsin Claude Code to confirm it loaded.

Bundled filesagents/openai.yaml

Details

From the project README

Find Improvements

Scope

Analyze the current repository unless another repository or narrower focus is selected. Start without parameter questions; report unavailable or ambiguous repositories rather than substituting.

Propose two bugs and two refactorings for the user's decision. Work read-only: do not edit files, install dependencies, switch checkout, implement fixes, create commits/PRs/tickets or change schedules.

Dependencies

No skill dependencies

Workflow

Analysis

  • Record start time, repository, branch and HEAD SHA. Read repository instructions, relevant architecture and tests. Preserve local changes; distinguish them from committed code.
  • Start with recent commits/diffs, prioritizing changes since a previous report's SHA if it is an ancestor of HEAD. Otherwise state the recent range reviewed. Continue into related code even without new commits.
  • Spawn exactly two subagents: one searches for up to two bugs, the other for up to two refactorings. Give both the scope, baseline, read-only constraints, candidate criteria and analysis deadline. Do not spawn one agent per finding or let these agents delegate further. If delegation is unavailable, report the limitation rather than silently replacing it with solo analysis.
  • Actively search and verify for at least ten elapsed minutes; parallel work does not shorten this requirement. Challenge candidates even if four appear early; never fill time with waits. Trace callers, contracts, state and data flow; prioritize regressions before broadening. Attribute findings to commits only with evidence.
  • Check accessible open PRs, previous reports and user decisions for duplicates, work in progress and rejected ideas; disclose unavailable sources. Validate subagent evidence, reject weak or duplicate findings and rank each category by value relative to effort. If blocked, report actual duration and cause.

Candidates

Bugs: concrete, reachable behavioral errors requiring an estimated 30–60 changed implementation lines each, excluding tests. Derive trigger, expected behavior and consequence from code and contracts. Exclude hypothetical risks, style preferences and unsupported product assumptions. Use existing tests or non-mutating reproductions where useful; distinguish execution from code analysis.

Refactorings: behavior-preserving improvements within roughly 200 changed implementation lines each, excluding tests. Explain the simplification, deduplication or structural change and concrete maintenance benefit. Exclude features, cosmetic changes and speculative abstractions.

Increase limits only with user authorization for that category. Estimate honestly; report fewer proposals if evidence or scope cannot support two. Do not invent findings.

Result

Respond here with repository, branch, HEAD SHA, reviewed range, elapsed time and coverage/verification limits. Retain the SHA for later runs without state files.

Use Bugs (B1/B2) and Refactors (R1/R2). For each proposal include:

  • Trigger and consequence, or proposed refactoring and benefit.
  • File links with line numbers, code-based justification, related commit if established, verification and a useful reproduction or regression check.
  • Estimated implementation lines excluding tests, and behavior to preserve.

State shortages. End with proposals for user selection; do not implement or ask for approval per item.

Comments

Markdown supported