Review agent work plans and track user intent logs
A developer logs every user message to distill intent and has agents review saved plan files against project constraints before executing work. ·
Done with a custom-built or unnamed agent
- What they did
- The commenter keeps a chat log of every user message so intent isn't lost when sessions are compacted or restarted, and periodically reviews it to distill the latest form of that intent. Before executing a stretch of work, they generate a plan, save it as a file, and have one or more agents check it against project constraints.
- How it went
- The source gives no results or numbers. The commenter frames both as practices for keeping code aligned with intent, and says the right way to build trust in agent code is project specific.
Try it yourself with Claude Code
Before you carry out any task in this project, write your plan to a file in [plans folder] and review it against these project constraints: [list constraints, e.g. no new dependencies, keep the API stable]. Also append each message I send to an intent log at [log file path]. Show me the reviewed plan and wait for my approval before you start changing anything, and once the work is done, summarize what my intent log says I have been trying to achieve.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: HN · Oct 29, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
One-shot a CLI coding agent on top of a Python LLM library
Simon Willison had the Fable model more or less one-shot a CLI coding agent built on his LLM Python library.
Custom agent
Rewrite a README after a rearchitecture
Devin scanned a rearchitected codebase and delivered a new README as a PR with accurate setup and API details, but struggled with design rationale and purpose.
Replace token-heavy Jira and GitLab MCPs with CLI skills
The author had Claude run --help on Atlassian and GitLab CLIs, saved the actions as Markdown, and built lightweight agent plugins that cut token use.
Have Copilot coding agent write and push code
A developer found the agent good at context and code, but it pushed to a branch leaving main empty and spent hours checking after being told to proceed.
Screen interview code submissions with a Codex review agent
A commenter uses Codex with a targeted agents.md for front-line code reviews, rejecting submissions with too many files changed or added.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.