AutoGineer
Generated banner for Skill Authoring
Generated logo for Skill Authoring

Skill Authoring

Grafana LabsOfficialPublished by Grafana Labs

  • Skill
  • Meta (authoring, orchestration, planning, memory)
  • Agent tooling
  • Apache-2.0
  • No issues found
Download ZIP

From the author

Author, audit, and improve Grafana SKILL.md files against Anthropic's published Agent Skills guidance and the four-dimension rubric the grafana/skills CI gate uses (conciseness, actionability, workflow clarity, progressive disclosure). Applies the canonical SKILL.md structure (YAML frontmatter + body + references/ + scripts/ + assets/), the "pushy description" trigger pattern, the three-level progressive-disclosure model, and the validate-fix-rerun feedback loop. Use when creating a new skill in this repo, when reviewing a skill PR, when a skill's Tessl review score is below 75 (the merge gate), when a skill's description isn't getting picked up by agents, when restructuring a long SKILL.md into a bundle, or when the user asks how to write, improve, optimize, audit, or fix a skill - even if they don't say "skill" explicitly (e.g. "this isn't triggering", "Tessl scored this 72", "split this doc").
— the description by Grafana Labs

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

This skill provides a structured approach to creating and maintaining agent skills within the Grafana ecosystem. It defines a canonical structure for SKILL.md files, including YAML frontmatter, progressive disclosure patterns, and a validation feedback loop. It is designed to help developers align their skills with specific rubrics covering conciseness, actionability, workflow clarity, and progressive disclosure, ensuring they pass automated CI gates and trigger effectively for users.

What it does

  • Provides a rubric for auditing and scoring skill documentation.
  • Defines the canonical SKILL.md structure and bundle layout.
  • Implements a validate-fix-rerun feedback loop for skill optimization.
  • Offers guidance on progressive disclosure and splitting strategies for large documentation.
  • Details trigger patterns to ensure skills are correctly identified by agents.
  • Lists common anti-patterns to avoid during skill development.

When to use it

  • When creating a new skill within the repository.
  • When reviewing a skill pull request.
  • When a skill fails the CI merge gate or receives a low review score.
  • When a skill is not triggering as expected during agent interactions.
  • When restructuring long documentation into a bundle.
  • When seeking advice on optimizing or fixing existing skill files.

Advantages

  • Standardizes skill development across the repository.
  • Provides actionable feedback based on automated review scores.
  • Reduces token usage through progressive disclosure strategies.
  • Ensures consistent trigger behavior via pushy description patterns.
  • Facilitates easier maintenance through modular bundle layouts.

Usage examples

“How do I improve my skill's review score?”
The agent will guide you through the Tessl review process, identifying specific dimensions like conciseness or actionability that need improvement based on the rubric.
“My skill isn't triggering when I ask for help. What should I do?”
The agent will audit your skill's description against the pushy-description pattern and suggest updates to ensure it includes the necessary trigger terms.
“How should I split a long SKILL.md file?”
The agent will provide strategies for splitting content into references/ files based on domain, depth, or format to maintain the three-level loading model.

Install

  1. Download the skill as a ZIP:

    https://autogineer.com/agent-skills/grafana/skill-authoring/download

  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip skill-authoring.zip -d .agents/skills

Documentation by Grafana Labs

Authoring & Improving Grafana Skills

How to write, review, and improve SKILL.md files so they pass the repo's CI gate and score well against the Anthropic-aligned rubric Tessl uses.

Critical rules (always)

  1. Description is the primary trigger — third-person, ≤1024 chars, must include explicit "Use when..." phrasing AND list concrete trigger terms users naturally say. See references/descriptions.md for the pushy-description pattern that combats undertriggering.

  2. Body under 500 lines — split into references/*.md if approaching the limit. SKILL.md is the routing layer, not the entire knowledge base.

  3. One level of nesting for references — link from SKILL.md directly, never SKILL.md → a.md → b.md. Claude may use head -100 previews on nested chains and miss content.

  4. Imperative voice — "Run X" not "You should run X" not "It is important to run X". Explain why over heavy-handed MUST markers.

  5. Concrete examples beat prose — copy-paste-ready commands, real config snippets. Tessl's actionability dimension scores this directly.

  6. No reserved words in name — anthropic and claude are forbidden in skill names.

  7. No time-sensitive language in the body — "after August 2025…" rots. Use an <details> "Old patterns" section for legacy info instead.

  8. Validate before committing — ./scripts/lint-skills.sh skills/<plugin>/<your-skill> clean + Tessl score ≥75 (run tessl skill review --json <dir>).

The rubric

CI fails any PR where a touched SKILL.md scores below 75 on four 0-3 dimensions: conciseness, actionability, workflow clarity, progressive disclosure. Full per-dimension scoring + Anthropic-doc mapping in references/rubric.md.

Score variance

The judge is an LLM and swings 7-10 points run-to-run. Local 94 commonly lands at CI 85. Ship only on three consecutive local 100s.

Decision tree for a new skill

  1. What product / domain does this skill belong to? Pick the right plugin folder: grafana-core/, grafana-cloud/, grafana-lgtm/, grafana-app-sdk/, grafana-k6/, grafana-plugins/. If none fits cleanly, ask the user before creating a new plugin group (a new group requires updating three marketplace.json files).

  2. Estimate body length.

  3. Write a "pushy" description first. The description is the only thing always loaded into context. If agents don't trigger the skill, nothing else matters. See references/descriptions.md for the pattern.

  4. Draft body with the four-dimension rubric in mind.

    • Cut every sentence Claude already knows (Conciseness)

    • Replace prose explanations with code blocks (Actionability)

    • Number every multi-step procedure + add a validation step at the end (Workflow clarity)

    • If you reach for <details>, consider whether that content belongs in references/ instead (Progressive disclosure)

  5. Register in marketplace manifests. Add the skill path to the skills array in all three:

    • .claude-plugin/marketplace.json

    • .cursor-plugin/marketplace.json

    • .agents-plugin/marketplace.json

  6. Validate locally.

    # 1. Lint clean (0 errors)
./scripts/lint-skills.sh skills/<plugin>/<your-skill>

# 2. Tessl reviewScore ≥75 (the CI gate)
tessl skill review --json skills/<plugin>/<your-skill> | jq '.review.reviewScore'

# 3. If below 75 or you want ≥85: run --optimize (requires auth)
tessl skill review --optimize --yes --max-iterations 3 skills/<plugin>/<your-skill>

    If the run fails: read the lint error / Tessl suggestion, fix, re-run. Don't open the PR until both checks pass cleanly. The feedback-loop pattern beats one-shot writing.

Fixing a low-scoring existing skill

  1. Read the judge's verbatim Suggestions text (non-JSON output):

    tessl skill review skills/<plugin>/<name>

    The Suggestions: block under each dimension names the exact sentences/sections to cut. Copy the suggestion — don't guess. Then verify the lowest dimension matches your read.

  2. Apply the fix pattern from references/rubric.md:

    • Conciseness 1-2 → cut intros, definitions, multi-line tables that mostly point to refs

    • Actionability 1-2 → replace prose with code blocks and CLI commands

    • Workflow clarity 1-2 → add numbered steps + validation checkpoints

    • Progressive disclosure 1-2 → split into references/*.md

  3. If the skill is intentionally a routing document (like grafana-k6/k6-docs), don't let --optimize inline the bundle back into SKILL.md. Hand-craft a minimal copy-paste "validation loop" inline so SKILL.md is independently actionable, while preserving the bundle.

  4. Re-score five times locally. Don't stop until all five runs hit 100 — see "Score variance" above for why.

Anti-patterns

See references/anti-patterns.md.

References

Files

references/anatomy.md

Anatomy of a skill bundle

The canonical layout

skills/<plugin>/<skill-name>/
├── SKILL.md              (required: YAML frontmatter + body)
├── references/           (markdown loaded only when needed)
│   ├── topic-a.md
│   └── topic-b.md
├── scripts/              (executable code; not loaded into context)
│   └── helper.py
└── assets/               (templates, fonts, images used as output)
    └── template.json

Source: anthropics/skills — skill-creator SKILL.md § Anatomy of a Skill.

The three-level loading model

LevelWhat's loadedWhenToken cost
1. Metadataname + descriptionAlways (at agent startup)~50-200 tokens per skill
2. SKILL.md bodyFull markdown bodyWhen the agent decides the skill triggers~500-3000 tokens once loaded
3. Bundlereferences/*.md, assets/*, scripts/* (via bash, not context)On demand, only the specific file the body links toZero until accessed

Implication: the cheapest way to lift Conciseness is to move heavy content from level 2 to level 3.

SKILL.md size guidance

  • Aim: ≤200 lines for the body

  • Hard guideline: ≤500 lines (Anthropic's documented limit)

  • Soft threshold: If you're past 250 lines, ask whether some sections should be references/*.md

Splitting strategies

By domain (when a skill spans multiple platforms/frameworks)
cloud-integrations/
├── SKILL.md           (overview + selection logic: "if AWS, see references/aws.md…")
└── references/
    ├── aws.md
    ├── gcp.md
    ├── azure.md
    └── confluent.md

Claude reads SKILL.md, sees the user mentioned AWS, reads only references/aws.md. The other three never load → zero token cost.

By depth (when overview vs. detail is clear)
react-19-plugin-migration/
├── SKILL.md           (decision tree + critical rules + happy path)
└── references/
    ├── eslint-9-migration.md  (deep dive for the ESLint sub-step)
    └── troubleshooting.md     (failure modes + fixes)
By format (when reference material doesn't fit the prose flow)
admin/
├── SKILL.md
└── references/
    ├── sso.md           (full OAuth/SAML/GitHub OAuth configs)
    ├── terraform.md     (Terraform provider examples)
    └── api-reference.md (Cloud API + Admin endpoints + audit logs)

Reference-file conventions

  • Filename: short, descriptive, lowercase-with-hyphens (troubleshooting.md, not Troubleshooting.md or Docs.md)

  • Top of file: if >100 lines, include a ## Contents table of contents so Claude can navigate during a partial read

  • Linked one level deep from SKILL.md — never a chain (SKILL.md → a.md → b.md)

  • Inline relative links — [full setup guide](references/sso.md) from SKILL.md, not absolute URLs

When NOT to inline

Tessl's rubric penalizes routing documents (a thin SKILL.md that defers to references/) by scoring them down for "not independently actionable". The naive response is to inline content from references back into SKILL.md, which undoes the deliberate progressive disclosure.

Better fix for routing skills: add a minimal copy-paste-ready "happy path" loop inline so SKILL.md is independently actionable for the most common task, while preserving the bundle for everything else.

Concrete example — grafana-k6/k6-docs (33 lines in its SKILL.md):

## Workflow

Pick the workflow that matches the user's intent:

- **Write documentation:** follow [references/workflows/write.md](references/workflows/write.md)
- **Review documentation:** follow [references/workflows/review.md](references/workflows/review.md)

## Validating examples (always run before committing)

Every code example in k6 documentation and release notes must execute cleanly against `k6@master`:

```bash
cd ~/path/to/k6
git checkout master
git pull
go run . run /path/to/script.js
```

If the run fails: read the error, fix the source, re-run. Don't commit a doc edit whose
example doesn't run clean. For the full multi-example workflow with parallel subagents,
see [references/testing-workflow.md](references/testing-workflow.md).

This pattern (the file lifted its score from 72 → 100):

  1. Keeps the routing layer + bundle architecture

  2. Adds enough inline content that Claude can do the common task without loading any bundle file

  3. Still defers full procedural detail to the references/

scripts/ directory

Use when the skill includes deterministic operations Claude shouldn't reinvent:

  • File transformations (PDF extraction, image resizing)

  • Validation harnesses

  • Output formatters

Scripts are executed via bash; their contents do not consume context tokens, only their output does. Big performance win for skills that run the same code repeatedly.

Make scripts executable (chmod +x) — the linter warns otherwise.

assets/ directory

Use for template files, fonts, images, or other static resources the skill produces as output. Like scripts, assets don't consume tokens until referenced.

Marketplace registration

After creating the skill directory, add it to the skills array in all three marketplace files:

  • .claude-plugin/marketplace.json

  • .cursor-plugin/marketplace.json

  • .agents-plugin/marketplace.json

The three must stay in sync. CI's validate job will fail any PR where they diverge.

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Security

No issues found by our static scan· No issues found

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Published by
Grafana Labs
Publisher
Grafana Labs (Official)
Source
grafana/skills · skills/grafana-core/skill-authoring · commit 8fb72a5
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by Grafana Labs or Grafana Labs.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download skill-authoring.zip