← ClaudeAtlas

agent-promptlisted

Prepare a grounded task or continuation prompt. Use when delegating fresh work or carrying progress, decisions, failures, and proof into a new agent session.
fmind/dot · ★ 4 · AI & Automation · score 77
Install: claude install-skill fmind/dot
# Agent Prompt Prepare instructions the receiving agent can use without this conversation. [prompt-design](../prompt-design/SKILL.md) owns prompts embedded in applications; native resume or compaction is preferable when it already preserves the needed state. ## Workflow 1. **Choose the mode**: fresh task, or continuation of work already underway; resolve the requested outcome, latest corrections, scope, and acceptance criteria. 1. **Ground the prompt**: inspect relevant repository instructions, source, tests, and installed dependencies; verify paths and commands and label assumptions. 1. **For a continuation**: check `git status --short --branch`, record completed proof, the exact stopping point, outstanding work, failed approaches, and reasons for material decisions. 1. **Draft** from [prompt-template.md](references/prompt-template.md); omit empty sections and add ordered slices only when execution needs them. 1. **Write and check**: resolve the root with `git rev-parse --show-toplevel`; save `.agents/prompts/<file>.md`, defaulting to `<YYYY-MM-DD>-<slug>.md`, or `~/.agents/prompts/` outside a repository. Verify the receiver can act without hidden context, then report the path. ## Gotchas - **Authority travels with the task**: distinguish authorized actions from proposed work; writing the prompt does not grant additional permission. - **Useful history**: preserve failures and proof gaps; cite code paths and lines instead of copying source or the shared persona. - **Work