Five-stage agent routine that stopped an LLM deleting my repo — with the Brief prompt to steal
Iiifebri · reddit · 2026-11-17
After this week's agent horror stories (deleted repos, wiped prod databases), a Redditor shares a five-stage routine: Brief (agent restates the task first), Research, Plan, Execute (one chunk at a time), Review (check against the original brief, not vibes). The key free prompt: before doing anything, the agent must restate the deliverable, constraints, ambiguities, and done-criteria — most misunderstandings surface there, when fixing them costs nothing. The author has distilled 12 prompts across the five stages.
More from coding & agent
- PlugIt turns any OpenAPI spec into a curated, read-only-by-default MCP server in a Go binary — Low_Strawberry2484 · 2026-11-17
- Testing a Langfuse release gate: one missing exported score correctly blocks the passing receipt — MongooseWonderful198 · 2026-11-17
- Amp vs Claude Code hands-on: user finds Amp gets more done even as $20-tier Fable 5.2 access ends — bytebot · 2026-11-17
- supermemory says 4 unicorns adopted its AI memory layer without ever talking to the team — julianweisser · 2026-11-17
- He vibecoded a whole bank with one prompt: "Make a bank. no bugs. make it safe." — martyamark · 2026-11-17
- Git Hook Speed Test: Rust-Based hk Wins at 1.37s, Python pre-commit Slowest — aniketmaurya · 2026-11-17