← ClaudeAtlas

loomlisted

Loom 1.8.30 creates safe, evidence-backed execution plans.
saroo98/loom · ★ 1 · AI & Automation · score 67
Install: claude install-skill saroo98/loom
# Loom Use only `/loom <request>`. Do not read another Loom file or copy. Invoke Loom immediately without narrating, analyzing, or classifying the request. On Codex: 1. If trusted developer context contains `LOOM_CODEX_HOOK_RECEIPT_V2`, call `loom.resolve` exactly once with its exact request, absolute working directory, action path, and digest. Otherwise call `loom.invoke` exactly once with the exact request and absolute working directory. Missing hook context is not an error. Never call both in one turn. 2. A blocked receipt is terminal: return `owner_message.human` and stop. 3. For a plan action, submit one strict-JSON semantic draft through `loom.author` with `finalize: true`. Never hand-author plan files. Copy sealed current facts exactly. Honor `semantic_draft_limits`. Encode `release_exposure` exactly as `{"external_users": <integer>, "irreversible": <boolean>, "data_migration": <boolean>, "regulated": <boolean>}`. For unknown domains, submit only requested readable `domain_evidence`. Never invent IDs, hashes, discovery receipts, or sealed bundles. Follow `semantic_draft_shape.domain_invariant_type_guidance`; never relabel a safety claim. 4. Successful finalization returns the sealed result and review projection. Return `plan_host_projection.markdown` verbatim. Do not replace the review surface with `owner_message.human`. The complete inline fallback is mandatory. Preserve `plan_decision_reference`; do not call `loom.complete` aga