Before You Ask AI for Help, Clean Up the Context

Most bad AI output does not begin with a bad prompt. It begins with a pile of material that looks useful because it is nearby: an old flyer, an informal note, a half-remembered commitment, a private detail, and one approved source buried in the middle.

The tempting move is to paste everything into a chatbot and ask for a clean answer. That often produces a clean-sounding answer. It does not make the underlying packet current, safe, or coherent.

Context Cleanup is a small practice for slowing down before that handoff. The goal is not to collect more material. It is to decide what the next task is allowed to treat as true.

Start with a low-stakes packet: perhaps notes for a fictional event. Find the source that is both current and authoritative. Then label the rest. One item may be stale. Another may conflict with the approved date. A third may be a useful tone example but not a source of facts. A fourth may contain private information that has no place in the work.

Those distinctions matter because a language model is good at smoothing. It may blend an old date with a current location, turn an unconfirmed idea into a call to action, or make an unnecessary personal detail sound relevant. The prose can be perfectly readable while the context is wrong.

The practical output of cleanup is a short brief. It says what the task is, who it is for, what facts are current, which constraints are active, what material is excluded, and what still needs a human answer. That brief is not glamorous. It is a boundary around later work.

You can use a chat tool after that, but only for a restatement check. Ask it to name the goal, source of truth, constraints, exclusions, and unresolved questions before it writes anything. If it quietly adds a registration link or assumes an access service has been confirmed, the check has done its job: it has shown you what needs correction.

The durable lesson is simple. More context is not automatically better context. The first useful action is often cleanup, not another request. A later tool may help with wording, structure, or options. It cannot responsibly decide which source is current or which unknown should remain visible.

Try it with an invented packet first. Keep the real work out of public tools until you know what is safe to share. Then carry the clean brief forward with its uncertainty intact.

Draft for review. Do not post externally until privacy, accessibility, and learning-design review are complete.

---

Source: labs/context-cleanup/companions/article-medium-draft.md