DBAdmin/inbox/skill_up.md

name, version, owner, description, trigger, inputs, outputs

name	version	owner	description	trigger	inputs	outputs
my-custom-skill	1.0.0	team-name / person	Performs a well-defined task related to [domain], producing a deterministic and auditable output.	manual cron condition	required optional name type description topic string Main subject provided by the user name type default verbosity enum(low|medium|high) medium	type markdown location memory / file / external-system

My Custom Skill

Purpose

This skill is responsible for one thing only:

[Clearly state the single responsibility]

It should be used when:

• ✅ The user intent matches [explicit criteria]
• ✅ All prerequisites are satisfied
• ❌ It must NOT be used when [explicit exclusion cases]

---

When to Use (Decision Logic)

Trigger this skill only if ALL conditions are met:

1. User intent is related to [topic/domain]
2. Required inputs are present and valid
3. No higher-priority skill is better suited

Fallback:

• If any condition fails → return control to the orchestrator

---

Prerequisites

Required

• Files:
  • /config/skill-config.yaml
• Secrets:
  • API_KEY_X
• Tools:
  • CLI: tool-x >= 1.2
  • Access to: [system/service]

Validation Checklist

• Config file exists
• Secrets resolved
• External dependency reachable

Abort execution if any check fails.

---

Execution Flow

Step 0: Pre-flight Validation (MANDATORY)

Goal: Fail fast, fail safe.

• Validate inputs schema
• Sanitize user-provided text
• Check permissions / access scope
• Log execution start with correlation ID

Output:

• Validation report (internal)

---

Step 1: Context Gathering

Goal: Build minimal, relevant context.

Actions:

• Load required files
• Query only necessary data
• Ignore unrelated information

Rules:

• No assumptions
• No hallucinations
• Prefer explicit data over inference

Artifacts:

• context.json

---

Step 2: Core Logic Execution

Goal: Perform the primary task.

Actions:

• Execute deterministic logic
• If using LLM:
  • Provide strict system instructions
  • Use constrained prompts
  • Avoid open-ended creativity unless explicitly required

Rules:

• One responsibility
• No side effects outside defined scope

Artifacts:

• result.raw

---

Step 3: Post-processing & Output

Goal: Produce clean, user-ready output.

Actions:

• Normalize formatting
• Remove internal metadata
• Apply verbosity level
• Validate final output

Output:

• User-facing result
• Storage:
  • Save to [location]
  • Notify [who/what] if applicable

---

Error Handling

Expected Errors

• Missing input → return actionable message
• External dependency unavailable → retry or abort gracefully

Unexpected Errors

• Log full context
• Return safe, non-technical message to user
• Escalate via monitoring

---

Observability & Auditing

Log at minimum:

• Skill name & version
• Trigger type
• Inputs (redacted)
• Execution time
• Outcome (success/failure)

Metrics:

• Success rate
• Avg execution time
• Most common failure reason

---

Common Mistakes

1. Skill doing too many things
2. Triggering on vague user intent
3. Missing validation step
4. Overusing LLM where deterministic logic is enough
5. No clear failure path

---

Example Use Case

User:

"Generate a summary of database performance issues from last week"

Skill Output:

• Structured markdown summary
• Saved to /reports/db-summary.md
• Notification sent to Slack #db-alerts

---

Notes for Future Improvements

• Add caching for repeated inputs
• Introduce dry-run mode
• Expand structured outputs (JSON schema)