菜单
Cluster near-duplicate entries in .claude/settings.json + settings.local.json allow list into broader patterns under a safety taxonomy. Propose bulk collapses (10 entries → 1 pattern), user approves per-cluster. Run standalone when the allowlist is getting bloated.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when finishing a development branch, after resolving a debugging session, or whenever a failure-then-fix cycle reveals engineering constraints worth preserving as a2o scenarios. Also invocable manually with /story-harvest.
Use when finishing a development branch, after resolving a debugging session, or whenever a failure-then-fix cycle reveals engineering constraints worth preserving as a2o scenarios. Also invocable manually with /story-harvest.
Use when starting a planning or brainstorm pass on a feature or subsystem (or grounding external prior-art) and you need current ground-truth across the architecture's relevant seams before designing — "we're exploring X, ground my understanding", "map the seams for X", or as the grounding step before /brainstorm, /shift, or app-port.
Use when reviewing, designing, debugging, or modifying any UI surface — Angular app pages, Lit elohim-elements, graphos Storybook stories (Library A default / Library B designed), the design guide — when judging visual quality, aesthetics, theming, or brand drift, or when a described feature can't be found in the rendered app. Especially when you're about to give a visual verdict from source files alone.
Mandatory gate for any feature design involving data entities (tables, models, routes, sync messages) OR identity/agency/role/capability framing. Forces P2P-native thinking — DHT entry types, content addressing, source-of-truth classification, and identity-ontology framing (imago-dei, not crypto self-sovereignty) — before proposing design approaches. Use when brainstorming any feature that creates, stores, references, or syncs data entities, or that names an identity/agency tier.
Orchestrate the four-agent memory team (librarian, historian, cartographer, storyteller) through a four-phase substrate-currency ceremony — population-wide drift triage, four-lens deep-read on 1-2 picked surfaces, storyteller-pen synthesis, apply + downstream coherence verification. The deliverable is substrate-grounded gospel-tier rewrites. Invoke when the substrate-currency audit accumulates drift signal, when a major substrate landing happens that the gospel-tier hasn't absorbed, or on operator request.
| name | generalize-permissions |
| description | Cluster near-duplicate entries in .claude/settings.json + settings.local.json allow list into broader patterns under a safety taxonomy. Propose bulk collapses (10 entries → 1 pattern), user approves per-cluster. Run standalone when the allowlist is getting bloated. |
| metadata | {"sourceRuntime":"claude","master":"package","governance":"epr:elohim-agent/skills/generalize-permissions"} |
Reduce Claude Code permission-prompt pain by replacing many literal
allowlist entries with fewer broader patterns. Safety is preserved via
the taxonomy in genesis/agentic/data/safety-taxonomy.json.
Run /generalize-permissions when the allowlist is large or you've
been approving near-duplicate entries frequently.
Load current allowlists.
Read:
.claude/settings.json (committed, durable — the permanent palette).claude/settings.local.json (local, shift-scoped additions)Run the generalization algorithm.
node --input-type=module -e "
import('./genesis/agentic/generalize.mjs').then(async ({ clusterAndPropose }) => {
const { readFileSync, existsSync } = await import('node:fs');
const tax = JSON.parse(readFileSync('genesis/agentic/data/safety-taxonomy.json', 'utf8'));
const durable = JSON.parse(readFileSync('.claude/settings.json', 'utf8')).permissions?.allow ?? [];
const local = existsSync('.claude/settings.local.json')
? JSON.parse(readFileSync('.claude/settings.local.json', 'utf8')).permissions?.allow ?? []
: [];
const all = [...durable, ...local];
const proposals = clusterAndPropose(all, tax);
console.log(JSON.stringify(proposals, null, 2));
});
"
Present each proposal to the user for review.
Group by safety tier. For each cluster, show:
Ask: "Apply this generalization? [y/n/skip]" per cluster. Bulk-approve groups where the user says "all broadly-safe proposals".
Apply approved proposals.
.claude/settings.json (durable),
unless it originated solely from settings.local.json shift-scoped
entries, in which case it stays local and tagged.Summarize and return.
Print: "<N> clusters applied, <M> literal entries collapsed to <P> patterns. Allowlist is now <X> entries (was <Y>)."
never_wildcard list in the taxonomy..bak with timestamp) before
writing modifications. If anything goes wrong, restore.Allowlist has fewer entries than when you started, every removal is justified by an approved generalization, and no novel patterns were introduced outside the safety taxonomy.