# Which Surface

> Decide whether a piece of work belongs in a skill, a subagent, a hook, CLAUDE.md, or none of them

- Skill: `ciscoittech/which-surface` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ciscoittech/which-surface`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ciscoittech/which-surface/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: ciscoittech (https://skillmd.com/u/ciscoittech)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/ciscoittech/which-surface

---


Decide where a piece of work belongs, given what you want to make repeatable.
Most of the time the answer is "nowhere new."

Frontmatter is restricted to the six Agent Skills spec fields on purpose, so this
skill uploads to claude.ai unchanged. `argument-hint` would be a hard error there.

Deliberately model-invocable: the moment this is useful is when someone asks
"should I make an agent for this?", which means the description has to be in
context to be matched. Contrast `/build-feature`, which sets
`disable-model-invocation: true` because firing it spawns agents.

## Ask in order. Take the first yes.

1. **Must this happen every time, even if the model is convinced otherwise?**
   → **Hook**, in `.claude/settings.json`.
2. **A short fact, true on every turn, wanted without anyone asking?**
   → **`CLAUDE.md`**. If it only applies to some files, a `.claude/rules/`
   file with a `paths:` field.
3. **Would doing this flood the conversation with output nobody rereads?**
   → **Subagent**.
4. **A procedure with known steps, started by name?**
   → **Skill**.
5. **None of the above?** → **Neither.**

## Then confirm the answer

Do not skip this. Each surface has one question that catches the common mistake:

- **Hook** — name the case where the model would be *right* to skip it. If you
  can, it is instructions, not a hook, and a hook here will fire on work where
  it makes no sense until someone disables it.
- **`CLAUDE.md`** — is it under three lines? If not, it is a rule or a skill.
- **Subagent** — write the delegation message now. It receives none of this
  conversation. If the message is getting long, the work is entangled: do it
  inline instead.
- **Skill** — will the body stay under ~150 lines? It sits in context for every
  turn after it loads. And if it is pure guidelines, do not add `context: fork`;
  a forked skill with no imperative task returns nothing and does not error.
- **Neither** — have you actually done this three times? Twice is a coincidence.

## Report

State: the surface, the ladder step that produced it, the exact file path to
create, and the one confirm-check it had to survive. If the answer is "neither",
say what to type instead.

Full reasoning: `learn/01-four-surfaces.md`. Cost model, the two-direction
composition table, and a worked example with two "neither" outcomes.
Details on any single key: `learn/reference/frontmatter-keys.md`.