Dev & Engineering code-review-fixesci-fixesminimal-diffsmall-diffsloop-engineeringcode-fixing

Minimal Fix Skill

Produce the smallest code change that fixes one explicit problem, never touching unrelated code.

FollowSkills review · FSRS-2.0
Not recommended
57/ 100 5-point scale 2.9 / 5
1 2 3 4 5 6
1Trust19 / 25 · 3.8/5

The skill explicitly requires respecting a path denylist, prohibits editing .env, auth, payments, and secrets, and provides an escalation path instead of violating editing. It also emphasizes worktree isolation to avoid untracked modifications and states that the verifier, not the agent, decides when work is done. These measures show attention to least privilege and human confirmation. However, the publisher is unverified, and specific data-flow transparency or rollback mechanisms are not detailed (only worktree isolation is mentioned). Therefore, deductions are made for missing granular evidence in static assessment.

2Reliability8 / 20 · 2.0/5

The instructions are coherent and consistent, with clear steps and a sample output format for the happy path. However, there is no built-in test suite or automated verification for this skill, and failure feedback mechanisms are not well described (e.g., how to handle failure to reproduce the issue). Also, handling of abnormal inputs (e.g., missing files, ambiguous issues) is not specified. Given the static calibration limit of 10, the evidence is insufficient to reach that cap, so a lower score is given.

3Adaptability10 / 15 · 3.3/5

The skill's scenario is clear, inputs and outputs are defined, trigger conditions are explicit ('only when the fix target is explicit'), and boundaries declare no drive-by refactors. However, it does not define non-fit ranges (e.g., ambiguous issues or multi-file changes). Environmental fit for Chinese users and mainland China network reachability is not assessed, so deductions are made.

4Convention11 / 15 · 3.7/5

The document is well-structured with clear headings, process, and output template, and includes rules. The repository has MIT license, but the skill lacks version number or changelog. Minimal examples and no troubleshooting section. Maintenance responsibility only via repository. Therefore, deductions on convention.

5Effectiveness6 / 15 · 2.0/5

Defined output is a fix proposal with summary and verification results, clear and directly usable. However, static review cannot verify actual correctness or completeness, and no real use cases or comparisons provided. Thus, a moderate-low score.

6Verifiability3 / 10 · 1.5/5

CI and test infrastructure exist but do not directly verify this skill's key paths. No third-party execution evidence or independent reproducible conclusions. Therefore, low score.

Evidence confidence:Low Reviewed Aug 07, 2026 Reviewed revision c57170dbf93f
The upstream repository has new commits since this review. The score still applies to the reviewed revision shown and may not cover the latest changes.
Before you use it
  • The skill depends on external services (e.g., GitHub) and project context, which may be restricted in mainland China; consider accessibility.
  • Static review did not verify the skill's effectiveness in real projects; test in a controlled environment first.
  • The publisher's identity is unverified; assess reputation before adoption.
Review evidence [1][2][3][4][5][6][7]
See the full review method →

What does this skill do, and when should you use it?

The Minimal Fix Skill guides an AI coding agent to fix exactly one well-scoped issue with the smallest diff that could work. It requires precise inputs such as an exact failure message, reviewer comment, or issue description, and it respects a path denylist (e.g. .env, auth/, payments/) along with project build/test commands. The process involves identifying the minimal root cause, making only necessary changes, running relevant checks, and outputting a structured summary. The skill is part of the loop-engineering monorepo, which also provides CLI tools like loop-init for scaffolding.

Reads the provided failure message, reviewer comment, or issue description as input. Optionally reads implicated file paths, build/test commands from AGENTS.md or project skills, and a path denylist from loop safety policy. Runs tests or lint commands to verify the change. Produces a Markdown report with target, diff summary, verification run, and risk assessment.

  1. A developer facing a CI failure wants the smallest fix without scope creep.
  2. An engineer addressing a code review comment wants to resolve only that point.
  3. A maintainer needs a quick, safe fix for a small urgent issue.
  4. An automated loop (e.g., CI sweeper) applies safe, well-scoped fixes unattended.

What are this skill's strengths and limitations?

Pros
  • Enforces minimal scope, preventing drive-by refactors.
  • Requires verification by running relevant tests/lint.
  • Outputs a structured summary for human review.
  • Enhances safety via path denylist and human gate.
Limitations
  • Only effective when the fix target is explicit; ambiguous issues need triage first.
  • Relies on external process to maintain denylist and build commands.
  • Lacks a built-in test suite; correctness depends on environment.
  • In unattended mode, there is still risk of unintended regressions if verification is insufficient.

How do you install this skill?

Clone or copy the repository: git clone https://github.com/cobusgreyling/loop-engineering.git. The skill is at skills/minimal-fix/SKILL.md. The repo is a monorepo with 41 skills and tools; to use just this skill, copy the skills/minimal-fix folder into your agent's skills directory. Alternatively, use the loop-init scaffolder (npx @cobusgreyling/loop-init .) which creates the necessary file structure for skills.

How do you use this skill?

Invoke as you would any agent skill, e.g., by providing inputs: the exact failure message or review comment, relevant file paths, and build commands. For example, prompt the agent: 'Use the Minimal Fix skill to fix this CI error: ... , file ... , build command ...'. The skill will guide the agent through the minimal fix process and output a structured proposal. Ensure AGENTS.md or project skills include build commands and the path denylist (from loop safety policy).

FAQ

Can this skill handle multiple issues at once?
The skill rules state one problem per invocation; multiple failures should escalate or triage first.
How does it handle security-sensitive paths?
It respects a path denylist from loop safety policy (e.g., .env, auth/, payments/) — if relevant, it escalates instead of editing.
Do I need extra tools?
The skill itself does not strictly require additional tools, but Git and Node.js are recommended to run tests; the repo also offers CLI tools for scaffolding and auditing.
Does the skill auto-merge changes?
No, it explicitly does not mark its own work done — a verifier decides. Human review is still determined by loop configuration.

More skills from this repository

All from cobusgreyling/loop-engineering

Related skills