AGENTS.md

1# prompts/skills — Agent Contract
2 
3## Scope
4 
5This subtree owns structured `SKILL.md` workflow prompts published as SDP protocol
6artifacts.
7 
8## Contract
9 
10Each skill defines an executable workflow: triggers, preconditions, steps, outputs,
11stop conditions, and recovery behavior.
12 
13Skills must keep YAML frontmatter valid and aligned with `sdp.manifest.yaml` while
14the manifest remains the inventory gate.
15 
16## Dependencies
17 
18Skills may reference root `AGENTS.md`, `docs/reference/skills.md`,
19`docs/reference/skill-authoring.md`, and task-specific reference docs.
20 
21Do not encode package API contracts here. Put subtree-specific facts in the
22nearest module-local `AGENTS.md`.
23 
24## Runtime Assumptions
25 
26Harnesses may load these files through Claude plugin format, symlinks, generated
27adapters, or explicit prompt paths. Keep prose harness-neutral unless a section is
28explicitly marked as harness-specific.
29 
30## Local Rules
31 
32- Keep workflows executable and bounded.
33- State refusal, blocker, and stop conditions explicitly.
34- State completion criteria explicitly: verification run, scoped staging, commit,
35  push or exact blocker.
36- Never instruct agents to use `git add .`; enumerate scoped files from the write
37  plan or command output.
38- Do not duplicate Beads, branch, or publish policy unless the skill changes how
39  that policy is applied.
40- Deprecated skills must route to the canonical intent or replacement skill.
41