cwijayasundara Author Prompt Surface
Claude-only discipline for the harness monorepo (and any plugin that ships agents/skills). Load before writing or rewriting a prompt surface.
Required reading
Read and follow docs/prompting-standards.md in full. That document is the
source of truth; this skill is the invocation trigger + checklist.
Surfaces this skill covers
| Path | Role |
|---|---|
.claude/agents/*.md |
Subagent prompts (model: frontmatter only for model pins) |
.claude/skills/*/SKILL.md |
Skill prompts and stage workflows |
.claude/commands/* |
Slash-command entrypoints |
Checklist (every new or edited prompt)
Copy from docs/prompting-standards.md — fail the edit if any box is open:
- No anti-laziness
CRITICAL/MUST/If in doubtleft in (true invariants only). - Every tool/subagent has a use this when … trigger condition.
- Finding/review steps say report everything with severity; gates filter downstream.
- Long-running steps audit progress claims against tool results.
- No "show/echo/transcribe your reasoning as text".
- Distinct blocks in XML tags; examples where behavior is subtle.
- Effort expectation noted for agentic/coding work (open
high/xhigh, then measure downward). - No instruction to verify / re-check / double-check its own work — that causes over-verification; structural verification (independent evaluator, deterministic gates) covers it.
- Nothing restated that the platform's base system prompt already installs (response length, scope discipline, tone, faithful reporting).
- Detail pushed to
references/*.md; the always-loaded surface carries only trigger + routing. - No model named in the prompt body; no directional nudge that assumes one model's default (criterion, not nudge).
Model pins
- Pin models only in agent frontmatter (
model:) or the session model. - Never name a model generation in the prompt body.
- One prompt body must run unchanged if the pin changes.
When not to use this skill
- Product application code under
src// generated apps. - Disposable artifacts (mockups, ARB narratives, research) — use the lite lanes.
- Editing deterministic hooks/scripts — use code-gen + tests, not prompting standards.
After editing
If the change is a new control (gate, sensor, reviewer), also follow
HARNESS.md / harness-manifest.json registration so the control is not orphaned.