Agent prompts work better when phrased as dependencies, not prohibitions

// related

• #0001

  #0001Agents & toolsr/ClaudeAI6 hours ago

  Learning App Building with `Claude Code`: Start from Architecture, Not Code

  50radar

  Claude CodeTerminal coding agent — edits and generates code from prompts

  Running code is not the finish line. Auth, APIs, hosting, version control, tests, analytics, and security decide whether a prototype can handle real users.

  • —The proposed stack map is simple: frontend, backend, database, and plumbing. The biggest skill gap sits in the plumbing layer, not screen generation.
  • —Claude Code can scaffold UI, backend, and database quickly, but it will not automatically make security, deployment, observability, or data boundaries solid.
  • —Good prompting should ask for architecture first: auth flow, API contracts, persistence model, testing path, and hosting assumptions before implementation.
  • —Best use case is turning app ideas into maintainable prototypes. Pure vibe coding burns tokens and hides production risks.

  Source: www.reddit.com/r/ClaudeAI/comments/1tonzj9/im_a_softwareRead original →
• #0002

  #0002Agents & toolsGitHub Changelog8 hours ago

  `Copilot Memory` adds deletion, repo scope controls, and CLI settings

  70radar

  Copilot MemoryCoding assistant memory — persists project context for suggestions

  Memory governance is getting less risky. Repo-level shutoff and CLI controls make preview trials easier.

  • —Improved deletion controls reduce stale or sensitive context risk when Copilot Memory learns the wrong thing.
  • —A repository-level off switch matters for client work, OSS repos, or codebases where persisted context is unwanted.
  • —More controls in the Copilot CLI move memory management closer to daily terminal workflows, not just web settings.

  Source: github.blog/changelog/2026-05-26-copilot-memory-has-moreRead original →
• #0003

  #0003Agents & toolsOpenAI Codex Releases10 hours ago

  `Codex CLI` `0.134.0` adds local history search and stronger MCP flows

  70radar

  Codex CLICoding agent CLI — automates code work from the terminal

  Local conversation search turns old agent work into reusable context. MCP, profile, and remote fixes make daily CLI use smoother enough to upgrade soon.

  • —Local history search supports case-insensitive content matches and previews, so prior fixes become searchable memory inside Codex.
  • —--profile is now the primary selector across CLI, TUI permissions, and sandbox flows. Legacy configs are rejected with migration guidance.
  • —MCP setup gains per-server environment targeting and OAuth options for streamable HTTP servers, reducing connector setup friction.
  • —Read-only MCP tools can run concurrently when readOnlyHint is advertised. Faster context gathering matters in tool-heavy agent runs.

  Source: github.com/openai/codex/releases/tag/rust-v0.134.0Read original →