ma-wiki-write

/ma-wiki-write

One skill to write your learned tier. You invoke it, then pick the operation by judgment — there is no menu to pre-select from. Carry the shared write discipline below into whichever operation fits.

Single-writer-per-page. Write only your own subtree. You never write another agent's subtree; cross-subtree refs you can only flag in your return for the lead.

Your write paths

<your-agent-name> is your literal agent filename without .md (e.g. ma-chain-decay-consultant), suffix and all. The slate path is absolute — never write a slate under a relative path, and never under .madagents/wiki/ (a slate written there is never auto-loaded and is silently inert).

Step 0 — should this be written at all? (the gates)

Two gates run before any operation. They are about whether to write, not which operation.

Signal gate (recording a finding). "Did I source-walk something non-obvious for THIS input — where a surprising default lives and how it behaves, a multi-step mechanism, a behavior pretraining wouldn't predict — such that a future dispatch will plausibly act better because of this page?" If yes, write — the source-walk is genuine learning worth caching. (Cache where the default lives and why it surprises, not the number — see discipline 4.) Don't pre-judge on "might future-me need it"; if it was non-obvious to walk, future-me benefits. Skip only when the walk was trivial restatement of something already loaded.

Record confirmations too, not only surprises and pitfalls. An approach you confirmed correct for THIS input — a path validated against source or probe that you'd otherwise only half-trust — earns a page, so the cache tells future-you what is safe to repeat, not only what to avoid. A tier holding only failures pushes the system toward over-caution.