Skill Authoring

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.

   • <200 lines of substance → single SKILL.md, no bundle

   • 200-500 lines → SKILL.md + references/<topic>.md for the long-form material

   • 500 lines → mandatory bundle split; see references/anatomy.md § Splitting strategies

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.

   Copy

   # 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):

   Copy

   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

• references/descriptions.md — the pushy-description pattern + trigger-term checklist

• references/rubric.md — per-dimension scoring with Anthropic-doc citations and concrete fix patterns

• references/anatomy.md — three-level progressive disclosure, bundle layout, splitting strategies

• references/anti-patterns.md — what NOT to do, with examples

• Anthropic — Agent Skills best practices

• anthropics/skills — skill-creator SKILL.md

• The Complete Guide to Building Skills for Claude (PDF)