Skip to main content
GLM models are reached through an Anthropic-compatible endpoint, so they run inside Claude Code against the same CLAUDE.md a Cordango workspace already ships.

Point Claude Code at GLM

Put that in ~/.claude/settings.json for every project, or in .claude/settings.json for this workspace only. Get a key from z.ai. The long timeout is not decoration. A whole-app generation pass is a long single call, and the default will cut it off partway. Everything on the Claude Code page applies from here: the same CLAUDE.md, the same permission allowlist, the same slash commands.

Write the sequence out

This is the one real difference in how you prompt. GLM follows an explicit, numbered plan closely and reliably. Left to infer the shape of the work, it tends to do more at once than you want, and a change spanning six aggregates is a change nobody can review. Say the steps.
Then ask for the screen change as its own turn.

Starting from nothing

What to watch for

Check that it actually ran check. Ask for the output rather than a summary of it. cordango check is free, so there is no reason for a step to be reported rather than run. Watch for a second copy of a core entity. The Phase 1 stop exists for this. A company or customer entity appearing in Phase 2 means Phase 1 was skimmed, and it is the one mistake that is genuinely awkward to unwind later. Keep the scope narrow. If a turn produced a diff touching more aggregates than you asked for, revert it and re-ask with the sequence written out. That is cheaper than reviewing it.
None of this is specific to GLM being worse. Explicit sequencing produces tighter diffs from every model on this list. GLM is just the one where the difference is most obvious.