| name | prompt-system-workflow |
| description | Change startup instructions safely. Triggers on prompt-system, AGENTS.md, CLAUDE.md, or instruction topology changes. |
| role | control-plane |
| argument-hint | [describe the instruction change] |
Prompt System Workflow
Purpose
Route startup-instruction changes to the right owner and prove health after edits.
Read First
docs/adr/0011-lean-startup-instructions.md
skills/prompt-system-router/SKILL.md
CONTEXT.md for vocabulary
Dependencies
skills/prompt-system-router/SKILL.md: hard dependency for classification.
scripts/agent-instructions.sh: hard dependency for delivery checks.
docs/adr/0011-lean-startup-instructions.md: owner-reference dependency.
- Missing router or health contract: blocked.
- Missing ADR: degraded; use
CONTEXT.md and startup checks, then report the missing ADR.
Resume State
- Treat canonical owner files, ADRs, and check output as durable state.
- Re-read
AGENTS.md, CLAUDE.md, router guidance, and touched owner files before resuming after compaction or a new session.
- Before stopping mid-change, record touched files, accepted decisions, remaining edits, and remaining checks in the owning tracker or handoff.
- Do not rely on transcript memory for which startup surface owns the rule.
Workflow
- Classify with router skill.
- Edit canonical owner:
- startup hot rules:
AGENTS.md
- Claude wrapper only if runtime proof requires it:
CLAUDE.md
- workflow depth: owning skill or context doc
- deterministic mechanics: code, CLI help, generated docs, or checks
- repo facts: repo-local
docs/agents/
- Never edit generated prompt artifacts.
- If delivery/check semantics changed, update
scripts/agent-instructions.sh.
- If decision trade-off changed, update or add ADR.
- Verify:
scripts/agent-instructions.sh check
bun scripts/multi-agent-smoke.ts --tests boundary,propagation
Use focused smoke only when shared startup behavior or delivery changed.