verification

Automating the bottleneck means finding the constraint that sets your pace and pointing Claude at it: four working rules distilled from Anthropic's own engineers.

Guardrails are the support structure that keeps an hours-long agent run honest: an audit instruction, a memory that survives resets, and fresh-eyed verifiers.

Wrong agent answers usually trace to the files the agent reads. The four ways a growing knowledge base drifts from reality, and the habits (a router file, read-only audits, scheduled pulls, backtracking) that catch the drift.

A rerun loop makes the agent redo its own work until a check passes. Script the repetitive half of a skill, end it with a pass or fail verdict, wrap both in a loop, and keep a human at the checkpoints where errors are expensive.

Audit Claude Code instructions, preserve project requirements, and test a smaller setup against the work you already trust.

Models got good where the work grades itself, and they aim at the most likely answer, which in design and writing is the average one. The fix: break your quality standard into checkable rules and keep human judgment for taste.

One big question in one chat makes the model decide, research, write, and grade itself in a single pass. Map the work instead: parallel research jobs, a checker that did not write the answer, a merge, and your approval before anything expensive.

One successful run proves an agent worked once. Collect real examples with answers a person already judged good, grade each run with a script or a judging model, and rerun the set after every change so you know whether it helped.

Turn a tested task into a reusable skill, choose where it runs, and catch failed work before it wastes more time.

Choose a task you need to do, learn from checked sources, then test your understanding by doing the work and explaining your decisions.

Compare agents on one repeatable task, check the tools each can use, and keep your skills usable when you switch.

Use failed runs to repair skill descriptions, scripts, and checks. Test each change before keeping it, with Harness Firmware as a working example.