@554/better Skill Creator

Better Skill Creator

You build skills that actually work — not just valid YAML, but skills that trigger reliably, complete their workflows without hand-holding, and compose well with other skills.

The fundamental insight from Anthropic's skill design guide: the description is the skill's interface to the agent. A skill that doesn't trigger is a skill that doesn't exist. A skill that triggers for the wrong things causes confusion. Get the description right first — everything else follows.

Phase 0: Use-Case Canvas (Do This Before Scaffolding)

Before writing a single line of SKILL.md, fill out this canvas. It forces clarity about what the skill actually does and prevents writing instructions for a skill that was never well-defined.

Skill Name: _______________

Use Case 1: _______________
  Trigger phrase: "..."
  Steps: 1. → 2. → 3.
  Expected output: _______________

Use Case 2: _______________
  Trigger phrase: "..."
  Steps: 1. → 2. → 3.
  Expected output: _______________

Dependencies:
  - API keys needed: _______________
  - Python packages: _______________
  - External services: _______________

Success looks like:
  - Agent triggers on ___% of relevant queries
  - User doesn't need to re-prompt to get _______________
  - Works independently when combined with: _______________

If you can't fill in at least 2 use cases with concrete trigger phrases, the skill isn't ready to build yet. Ask the user to clarify.

Phase 1: Understand the Request

• What capability? API integration, workflow automation, knowledge domain, meta-tool?
• What triggers it? Map the exact phrases a user would say. This becomes the description.
• What freedom level? Can the agent improvise the steps, or must they be exact?
  • High freedom → natural language guidance in body
  • Medium freedom → pseudocode + key params
  • Low freedom → scripts in scripts/ executed via bash
• What does success look like? Define this now, before you build — see Phase 4.

Phase 2: Scaffold

python skills/skill-creator/scripts/init_skill.py my-new-skill --path ./skills

With resource directories:

python skills/skill-creator/scripts/init_skill.py api-helper --path ./skills --resources scripts,references

Phase 3: Write the SKILL.md

3a. Craft the Description (Most Important)

The description is the only part of your skill that's always in context. It's what the agent reads to decide whether to activate the skill. Two requirements:

1. What it does — one clause
2. When to trigger — specific scenarios, not vague categories

Formula: [What it does] + [Use when <specific trigger conditions>]

Examples:

# Too vague — agent won't know when to use this
"Trading utilities and market analysis tools."

# Good — clear what/when
"Fetch live and historical crypto prices, on-chain metrics, and sentiment
scores. Use when the user asks about price, volume, market cap, trend, or
wants data to support a trade thesis."

# Good — multi-scenario trigger
"Generate TradingView-style candlestick charts with indicators. Use when
the user wants a visual chart, price visualization, or technical analysis
plot."

Hard limits:

• Max 1024 characters
• No XML tags < > — they break the skill loader
• Do not name the skill claude, anthropic, or any Anthropic product name

Trigger failure modes:

• Skill never activates → description too vague or missing "Use when"
• Skill activates for unrelated requests → description too broad
• Skill competes with another skill → descriptions overlap; narrow the trigger

3b. Frontmatter Template

---
name: skill-name            # lowercase, alphanumeric + hyphens, 2-64 chars
version: 1.0.0              # semver — required for market publishing
description: "..."          # trigger text — see formula above
author: your-name           # for market publishing
tags: [tag1, tag2]          # for discoverability

metadata:
  starchild:
    emoji: "🔧"
    skillKey: skill-name
    requires:
      env: [API_KEY_NAME]   # env vars that must be set
      bins: [python3]       # binaries that must ALL exist
      anyBins: [curl, wget] # binaries where ANY ONE is sufficient
    install:
      - kind: pip
        package: pandas
      - kind: apt
        package: curl
        bins: [curl]

user-invocable: true
---

3c. Body Structure — Choose a Pattern

Match the body structure to the skill's nature. Don't force workflows onto knowledge skills or vice versa.

Write for the agent, not the user. Direct voice.
Bad: "This skill can be used to generate charts when the user requests them."
Good: "You generate charts. Fetch OHLCV data, configure the chart layout, render via the charting script, and serve with preview_serve."

What to include:

• Domain knowledge the agent doesn't have from tool descriptions
• Decision trees: "when X use Y, when Z use W"
• Gotchas and edge cases
• Key thresholds and parameters

What to leave out:

• Tool parameter docs (already in system prompt)
• Generic programming advice
• README / changelog content
• Step-by-step for trivially obvious sequences

3d. On-Chain Skills — Wallet Policy Prerequisite

If the skill involves any on-chain operations (transactions, approvals, swaps, bridging, signing):

## Prerequisites — Wallet Policy

Before executing any [operation], the wallet policy must be active.
Load the **wallet-policy** skill and propose the standard wildcard
policy (deny key export + allow `*`). This covers all [skill-name]
operations across all chains.

3e. Composability Check

Before finalizing: does this skill assume it's the only capability active? It shouldn't.

• Does it use tool names that might conflict with other skills? Rename.
• Does the description overlap with an existing skill? Narrow the trigger.
• Does it hardcode assumptions about what other skills exist? Remove them.
• Would it break if loaded alongside the skill it's replacing? Test it.

Phase 4: Define Success Criteria

Define these before testing, not after. Write them into the skill's comments or your notes.

Quantitative (aim for these thresholds):

• Triggers on ≥90% of queries you listed in the Use-Case Canvas
• Does NOT trigger on >5% of unrelated queries
• Completes its workflow in ≤ the number of tool calls you expect
• Zero API call failures due to wrong parameters

Qualitative:

• User doesn't need to re-prompt or correct the agent mid-workflow
• Output is consistent across sessions with the same input
• Works alongside other skills without conflicts

Phase 5: Validate (Syntax + Behavior)

Syntax validation

python skills/skill-creator/scripts/validate_skill.py ./skills/my-new-skill

This checks: frontmatter structure, required fields, TODO placeholders, body length, field placement.

Behavioral testing

Run 3+ test cases from your Use-Case Canvas and evaluate against your success criteria:

Test: "[exact user phrase from canvas]"
Expected: skill triggers, completes workflow X
Actual: _______________
Pass/Fail: ___

Common failure modes and fixes:

Edge cases to test:

• Empty/ambiguous user input
• Multi-step workflows where earlier steps fail
• Concurrent use with another active skill
• First use (no prior context) vs. mid-conversation use

Phase 6: Refresh and Confirm

skill_refresh()

Then run one final test from your Use-Case Canvas to confirm the skill loads, triggers, and completes.

Anatomy of a Well-Structured Skill

my-skill/
├── SKILL.md          # Frontmatter + core instructions (< 500 lines)
├── scripts/          # Fragile operations: exact API calls, rendering pipelines
│   └── render.py     #   Executed via bash, never loaded into context
├── references/       # Detailed docs loaded on demand
│   └── api-guide.md  #   Only read when agent explicitly needs it
└── assets/           # Templates, images, config files used in output
    └── template.json #   Never loaded into context

Context budget:

• SKILL.md body: loaded on every activation — keep it lean, ≤ 500 lines
• scripts/: executed, not read — no context cost
• references/: read on demand — only when explicitly needed
• assets/: never in context — output artifacts only

What NOT to Include

• README.md — SKILL.md IS the readme
• CHANGELOG.md — skills aren't versioned packages
• Generic best practices — "use error handling" is noise, specific gotchas are signal
• Tool parameter docs — already in the system prompt
• Security credentials — never hardcode API keys; use requires.env and env vars

Publishing to the Skill Market

When the skill is ready to share, publish it via the skill-installer workflow.

Pre-publish checklist:

• name field: lowercase hyphen-case, 2-64 chars
• version field: semver (e.g. 1.0.0) — immutable once published
• description passes the formula test (what + when)
• author and tags set
• No [TODO] placeholders in body
• Validate passes with no errors
• Behavioral tests pass for all Use-Case Canvas scenarios
• No XML tags < > in description

Publish:

# 1. Get OIDC token
TOKEN=$(curl -s --unix-socket /.fly/api \
  -X POST -H "Content-Type: application/json" \
  "http://localhost/v1/tokens/oidc" \
  -d '{"aud": "skills-market-gateway"}')

# 2. Build and publish
SKILL_DIR="./skills/my-skill"
GATEWAY="https://skills-market-gateway.fly.dev"

PAYLOAD=$(python3 -c "
import os, json
files = {}
for root, dirs, fnames in os.walk('$SKILL_DIR'):
    for f in fnames:
        full = os.path.join(root, f)
        rel = os.path.relpath(full, '$SKILL_DIR')
        with open(full) as fh:
            files[rel] = fh.read()
print(json.dumps({'files': files}))
")

curl -s -X POST "$GATEWAY/skills/publish" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d "$PAYLOAD" | python3 -m json.tool

Version is immutable after publishing — bump to 1.0.1 before republishing.