skill-improve
/skill-improve
Evaluate and improve existing agent skills using observed failures and realistic task scenarios. Use for auditing skill behavior and instructions rather than creating a new skill from scratch.
- Author
- dnlrsls ↗
- Repository
- gentle-stack ↗
- Licence
- MIT
- Last repository change
- 03.11.2026
Assessment and security
- Files in this package
- 1
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 03.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Evaluate and improve existing agent skills using observed failures and realistic task scenarios. Use for auditing skill behavior and instructions rather than creating a new skill from scratch.
Original author description
Evaluate and improve existing agent skills using observed failures and realistic task scenarios. Use for auditing skill behavior and instructions rather than creating a new skill from scratch.
How it works
- Establish the problem
- Evaluate
- Improve and check
- Result and memory
SKILL.md preview
--- name: skill-improve description: Evaluate and improve existing agent skills using observed failures and realistic task scenarios. Use for auditing skill behavior and instructions rather than creating a new skill from scratch. --- # Skill Improve Improve a skill's decisions and observable outcomes while preserving its intended scope. More instructions are useful only when they solve a demonstrated problem. ## Establish the problem Identify the target skill, its intended consumers, the requested mode (audit or edit), and the relevant version. Read its entrypoint and only the supporting resources used by the workflow. Retrieve relevant Engram findings if they are not already in the current context. Use reported runs, artifacts, and user expectations to distinguish an instruction defect from a missing tool, host limitation, or ambiguous product decision. Do not promise that prose can enforce isolation, permissions, persistence, or reviewer independence. Inspect discovery and routing as well as execution: a vague description can select the skill for unrelated work, while a narrow one can miss its actual task. Preserve agent portability unless the skill explicitly targets a particular host. ## Evaluate - Choose representative cases, including the reported failure and an adjacent case where the proposed restriction should not apply. Define observable acceptance conditions before changing instructions. - Reproduce the failure in a disposable workspace when practical. Use drafts or a sandbox for external mutations; evaluation does not authorize publishing tickets, spending on additional agents, or changing real project configuration. - Separate structural validation, manual instruction walkthroughs, actual task execution, and independent evaluation. Do not claim a scenario passed because the text mentions the expected rule. - Use a separate evaluator only when supported and authorized. Give it the request, skill version, and necessary raw artifacts without supplying the intended answer. If independence is unavailable, label the evaluation author-directed. ## Improve and check For an audit-only request, return findings and proposed edits. For authorized changes, make the smallest change supported by evidence. Prefer removing duplication, clarifying a decision boundary, or adding a focused reference over introducing universal gates and new files. Keep skill names, descriptions, and resource paths consistent. Preserve standalone installation: a sibling skill can be an optional collaborator, never an undeclared required file. Check affected links and frontmatter with available validators; execute changed scripts. Repeat behavioral checks relevant to the change, retaining unaffected evidence with its original scope. Avoid wording-matching tests that merely echo the instructions. Stop when the requested defect is addressed, remaining findings are documented, or a concrete blocker prevents evaluation. Do not automatically modify unrelated skills, installed copies, or host configuration. ## Result and memory Return the original problem, changes or audit findings, evidence before and after, the skill version examined, and untested cases. Publishing or installing changes follows the user's existing authorization. Save durable failure patterns, decisions, and evaluation references in local Engram. In composed work, return them to the coordinating memory owner rather than saving a duplicate checkpoint. Resolve the actual project and use only a bound or successfully registered session if required. Read back uncertain saves before retrying, handle returned conflicts through supported operations, and disclose unavailable persistence. Never enable cloud to repair local memory.
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · dnlrsls/gentle-stack | 100 | MIT |
| yrzhe/claude-skills | 100 | MIT |
| chatquill/claude-code-websites | 90 | MIT |
Database snapshot: 13.11.2026 · package version 58f5ad1ea0d2 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100evaluate-skill
Evaluate an Agent Skill's routing, execution quality, reliability, composition, and incremental value with isolated realistic cases and evidence. Use when deciding readiness, investigating missed or false triggers, comparing skill versions, or hardening a skill after observed failures. - 100skill-generator
Generate new agent skills from scratch and update, refactor, or improve existing skills the user points to. Covers interviewing the user, drafting SKILL.md with bundled scripts/references/assets, auditing an existing skill against best practices, applying targeted edits, running… - 100eval-skills
Blind-test and improve an agent skill against realistic golden cases. Use when the user asks to evaluate, harden, or improve a skill; reports that a skill repeatedly misses an outcome; or provides a skill with concrete input and quality-bar examples. - 100skill-creator
Create, edit, improve, or audit skills for SwarmClaw agents. Use when creating a new skill from scratch or when asked to improve, review, audit, tidy up, or clean up an existing skill or SKILL.md file. Also use when editing or restructuring a skill directory. Triggers on phrases… - 90skill-lifecycle
Create, evaluate, improve, and benchmark content skills using the local Skill Lab workflow. Use when adding a new skill, tuning an existing skill's trigger behavior, iterating on SKILL.md instructions, or deciding whether a candidate skill should replace the current version. - 100skill-creator
Create new skills or improve existing skills by clarifying intent, drafting SKILL.md instructions, organizing supporting resources, and doing lightweight manual review. Use when users want to create a skill from scratch, revise a skill, package a skill, or improve a skill's trigger description.