# Action vocabulary

Skill bodies describe behavior with a small set of vendor-neutral verbs instead of naming a
specific coding agent or product. This keeps every skill usable by any agent that can read
Markdown and follow instructions. Repository maintenance checks guard against vendor names
creeping into skill bodies; the installed contract remains provider-neutral.

An **Action** is one bounded operation, not a synonym for a Skill, Tool, or Hook. See the
[canonical vocabulary](canonical-vocabulary.md) for the broader harness taxonomy.

Each verb below is defined by three facets: **Intent** (why a skill would use it), **Authority**
(what it is and is not allowed to change), and **Output** (what it leaves behind for the next
step).

- **Read**
  - Intent: gather existing information (a file, config value, or prior artifact) before
    acting on it.
  - Authority: read-only; never creates, edits, or deletes anything.
  - Output: loaded content available to inform the next step (for example, reading
    `.sdd-agentic-flow/config.yml` before using its artifact paths).
- **Write**
  - Intent: record a decision or result as a durable artifact.
  - Authority: create or update exactly the file(s) the current step is scoped to; never touch
    unrelated files.
  - Output: a new or updated file on disk (for example, `spec.md` with requirements and
    acceptance criteria).
- **Delegate**
  - Intent: hand a bounded sub-task to another skill or a human without doing the work directly.
  - Authority: names the receiving skill or human decision-maker; never performs the delegated
    work itself.
  - Output: a clear handoff (for example, delegating the PR write-up to `saf-create-pr` once
    evidence is captured).
- **Inspect**
  - Intent: examine code, tests, or evidence to understand current behavior.
  - Authority: read-only; never changes the inspected material.
  - Output: findings that inform a later Write, Verify, or Review step (for example, inspecting
    changed files for scope drift before classifying a task).
- **Verify**
  - Intent: check that a specific claim or artifact meets a defined gate.
  - Authority: read-only against the target, though it may run an approved local command to
    obtain evidence.
  - Output: a pass/fail determination backed by evidence (for example, confirming a RED command
    actually failed before recording it).
- **Review**
  - Intent: assess a completed artifact for correctness, safety, or quality.
  - Authority: read-only against the artifact; raises findings rather than fixing them directly.
  - Output: findings compared against defined acceptance criteria (for example, reviewing a
    diff against a task's acceptance criteria).
- **Ask Human**
  - Intent: pause for an explicit authorization or a missing decision only a human can make.
  - Authority: cannot proceed past the pause point without an explicit human answer.
  - Output: a recorded decision or authorization that unblocks the next step (for example,
    asking before creating `.sdd-agentic-flow/context/domain-glossary.md`).
- **Summarize**
  - Intent: condense findings, evidence, or a decision into a short report.
  - Authority: read-only against the source material; adds no new claims beyond what was
    gathered.
  - Output: a short report or handoff note (for example, summarizing unresolved decisions
    before handing off the spec package).