name

skill-author

description

How to author, improve, and trigger-tune skills for DaxAlgo Terminal's .claude/skills/ library — frontmatter contract, "pushy" descriptions that actually trigger, the bespoke-vs-external decision (and the licensing rule), and how to wire a new skill to the agents that should load it. Use whenever adding a new skill, fixing one that won't trigger, or deciding whether to vendor an external skill. Make sure to use this skill whenever the user mentions writing/adding/editing a skill, a SKILL.md, agent capabilities, or "teach Claude to do X repeatably."

Skill author — how we grow the skill library

The recipe for adding a skill to DaxAlgo Terminal. Clean-room adaptation of the public skill-authoring workflow (credit: the skill-creator skill in anthropics/skills), written in our own words and tailored to this repo's conventions. No external skill is vendored verbatim — see "External skills" for why and how.

A skill is a folder + a SKILL.md

.claude/skills/<kebab-name>/
└── SKILL.md          # YAML frontmatter (name, description) + markdown body
    (+ optional bundled scripts/reference files the skill points to)

name — kebab-case, matches the folder.
description — the only trigger signal. Put what it does AND when to use it here, not in the body. Claude tends to under-trigger skills, so make the description a little pushy: name the concrete files/phrases/contexts that should fire it. Compare:
- weak: "Conventions for the market-data pipeline."
- strong: "…Use when adding store tables, changing ingest normalization, wiring trade-tape into a broker, or debugging 'no data'/'duplicate ticks'/'wrong timestamps', or touching anything under src/TradingTerminal.MarketData/."
Body — the recipe: terse, imperative, example-driven. The why only where non-obvious.

House style for DaxAlgo skills

Match the existing skills (navigator, add-strategy, quant-math, …):

Lead with a one-line statement of what the skill is for, then the recipe.
Encode the layer-graph constraints the work touches — a skill that doesn't mention the graph where it's relevant is incomplete.
Reference real seams/paths (IBrokerClient, LiveSignalStrategyViewModelBase, Infrastructure/Backtest/Strategies/) so the worker doesn't guess.
Keep it ASCII-safe if any bundled script is PowerShell (Windows 5.1 mangles non-ASCII in no-BOM files — see .claude/hooks/verify-on-stop.ps1).
End with "what NOT to do" when there are sharp edges.

Wire it up (a skill nobody loads is dead weight)

Add the skill folder + SKILL.md.
Add a row to the agent → skill table in .claude/MULTI-AGENT.md.
Add a ## Load first line to each owning agent in .claude/agents/<agent>.md so the worker loads it automatically (that's how quant-math reaches the strategies/tool-windows agents).
If it's broadly useful, mention it in the CLAUDE.md skills table.

Does it trigger? (cheap eval)

You don't need a benchmark harness — sanity-check triggering by paraphrase:

Write 3–5 prompts a user would actually type for this skill's job.
For each, ask "does the description make this skill the obvious match?" If a prompt wouldn't fire it, the description is too narrow — widen it with the missing phrase/context.
Watch for over-trigger too: if it would fire on unrelated prompts, tighten the "Skip when…".

Bespoke vs external — and the licensing rule

Bespoke (default): author it here, tailored to the repo. Everything in .claude/skills/ today is bespoke.
External: useful skills exist in the wild (anthropics/skills, VoltAgent/awesome-agent-skills, plugin marketplaces via /plugin). Before vendoring any of them into this repo, check the license. If it's MIT/Apache/CC-BY, you may copy with attribution in the SKILL.md. If there's no license (e.g. anthropics/skills is license: null = all-rights-reserved) or a copyleft one (AGPL/GPL), do NOT copy the file — clean-room it: read it for the ideas, then write our own in our own words and credit the source (the same rule this repo applies to the Fincept review). This skill itself is an example of that.

How to pull an external skill (mechanism)

# Browse/install plugin bundles (skills + agents) from a marketplace:
#   in Claude Code:  /plugin   -> add marketplace -> install
# Or vendor a single license-clean skill by hand:
git clone --depth 1 <repo> $env:TEMP\ext-skill
#   read its LICENSE; if permissive, copy the one skill folder + keep its attribution:
Copy-Item $env:TEMP\ext-skill\skills\<name> .claude\skills\<name> -Recurse
#   then add the agent->skill wiring above. If NOT license-clean: clean-room instead.

What NOT to do

Don't put "when to use" guidance in the body — it must live in description to trigger.
Don't ship a skill without wiring it to an agent or the CLAUDE.md table.
Don't vendor unlicensed/copyleft skills — clean-room and credit.
Don't write a 500-line skill; if it's that big, it's probably two skills.

Skill author — how we grow the skill library

A skill is a folder + a SKILL.md

.claude/skills/<kebab-name>/ └── SKILL.md # YAML frontmatter (name, description) + markdown body (+ optional bundled scripts/reference files the skill points to)

name — kebab-case, matches the folder.

description — the only trigger signal. Put what it does AND when to use it here, not in the body. Claude tends to under-trigger skills, so make the description a little pushy: name the concrete files/phrases/contexts that should fire it. Compare:

weak: "Conventions for the market-data pipeline."
strong: "…Use when adding store tables, changing ingest normalization, wiring trade-tape into a broker, or debugging 'no data'/'duplicate ticks'/'wrong timestamps', or touching anything under src/TradingTerminal.MarketData/."

Body — the recipe: terse, imperative, example-driven. The why only where non-obvious.

House style for DaxAlgo skills

Match the existing skills (navigator, add-strategy, quant-math, …):

Lead with a one-line statement of what the skill is for, then the recipe.

Encode the layer-graph constraints the work touches — a skill that doesn't mention the graph where it's relevant is incomplete.

Reference real seams/paths (IBrokerClient, LiveSignalStrategyViewModelBase, Infrastructure/Backtest/Strategies/) so the worker doesn't guess.

Keep it ASCII-safe if any bundled script is PowerShell (Windows 5.1 mangles non-ASCII in no-BOM files — see .claude/hooks/verify-on-stop.ps1).

End with "what NOT to do" when there are sharp edges.

Wire it up (a skill nobody loads is dead weight)

Add the skill folder + SKILL.md.

Add a row to the agent → skill table in .claude/MULTI-AGENT.md.

Add a ## Load first line to each owning agent in .claude/agents/<agent>.md so the worker loads it automatically (that's how quant-math reaches the strategies/tool-windows agents).

If it's broadly useful, mention it in the CLAUDE.md skills table.

Does it trigger? (cheap eval)

You don't need a benchmark harness — sanity-check triggering by paraphrase:

Write 3–5 prompts a user would actually type for this skill's job.

For each, ask "does the description make this skill the obvious match?" If a prompt wouldn't fire it, the description is too narrow — widen it with the missing phrase/context.

Watch for over-trigger too: if it would fire on unrelated prompts, tighten the "Skip when…".

Bespoke vs external — and the licensing rule

Bespoke (default): author it here, tailored to the repo. Everything in .claude/skills/ today is bespoke.

External: useful skills exist in the wild (anthropics/skills, VoltAgent/awesome-agent-skills, plugin marketplaces via /plugin). Before vendoring any of them into this repo, check the license. If it's MIT/Apache/CC-BY, you may copy with attribution in the SKILL.md. If there's no license (e.g. anthropics/skills is license: null = all-rights-reserved) or a copyleft one (AGPL/GPL), do NOT copy the file — clean-room it: read it for the ideas, then write our own in our own words and credit the source (the same rule this repo applies to the Fincept review). This skill itself is an example of that.

How to pull an external skill (mechanism)

# Browse/install plugin bundles (skills + agents) from a marketplace: # in Claude Code: /plugin -> add marketplace -> install # Or vendor a single license-clean skill by hand: git clone --depth 1 <repo> $env:TEMP\ext-skill # read its LICENSE; if permissive, copy the one skill folder + keep its attribution: Copy-Item $env:TEMP\ext-skill\skills\<name> .claude\skills\<name> -Recurse # then add the agent->skill wiring above. If NOT license-clean: clean-room instead.

What NOT to do

Don't put "when to use" guidance in the body — it must live in description to trigger.

Don't ship a skill without wiring it to an agent or the CLAUDE.md table.

Don't vendor unlicensed/copyleft skills — clean-room and credit.

Don't write a 500-line skill; if it's that big, it's probably two skills.