菜单
Use when the user asks for ultracode, dynamic workflows, Claude Code /workflows-style orchestration, subagent swarms, parallel agents, multi-agent fan-out, broad migrations, codebase-wide audits, high-risk plans, or complex work that needs independent verification.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/dedene/skills --skill ultracode命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
Fetch TikTok transcripts/captions into cached local files, for a single video or all videos of a channel/user, falling back to local mlx-whisper transcription when a video has no native captions. Use when the user provides a TikTok URL, video ID, or @handle and asks to get, fetch, download, search, excerpt, summarize, or otherwise use TikTok transcripts/captions/subtitles.
Delegate a bounded, well-specified coding task to GPT-5.5 via the Codex CLI (codex exec) and verify the result. Use for bulk or mechanical work routed to gpt-5.5 per the model policy — clear-spec implementation, migrations, repetitive refactors, test backfills, data transforms — or when the user says "ask codex", "let codex build/fix this", "delegate to gpt-5.5", or "codex exec". Not for taste-critical UI, copy, or public API design.
Independent GPT-5.5 code review via the Codex CLI — codex review for uncommitted changes, a branch diff, or a commit; adversarial exec mode for challenging a design or premise. Use when the user asks for a codex or gpt review, a second opinion on a diff or plan, or when adding gpt-5.5 as an extra review perspective alongside Claude reviews. Verifies findings against the code before reporting; never auto-fixes.
| name | ultracode |
| description | Use when the user asks for ultracode, dynamic workflows, Claude Code /workflows-style orchestration, subagent swarms, parallel agents, multi-agent fan-out, broad migrations, codebase-wide audits, high-risk plans, or complex work that needs independent verification. |
Ultracode is the Codex approximation of Claude Code dynamic workflows. It should feel like a workflow run, not like a longer chat: define phases, fan out bounded agents where useful, keep intermediate state outside the main conversation, show progress visibly, reduce and cross-check results, then return one coordinated answer.
Mimic the functional shape of dynamic workflows:
Do not mimic dynamic workflows by adding ceremony to small work. If a normal Codex pass is faster and safer, use Direct mode and say why.
Claude Code dynamic workflows move orchestration into a script and show it through /workflows: phases, agent counts, token/tool usage, elapsed time, pause/resume/stop/restart controls, drill-down into agents, and optional saved commands.
Codex does not have that native runner unless a concrete runner tool is present. In Codex, Ultracode emulates the behavior with:
multi_agent_v1.spawn_agent workers when available and the user has opted into workflow/subagent workmulti_tool_use.parallel for independent local reads and command checksupdate_plan as the live task panel.workflow/ultracode/<run-id>/ as the durable run ledger when resumability or auditability is worth the costNever pretend the main thread is multiple agents. If no real worker primitive exists, use Direct or Thin mode rather than roleplaying a workflow.
Choose the lightest mode that preserves the workflow value.
| Mode | Use when | Mechanics |
|---|---|---|
| Direct | Small or tightly coupled work; one agent can finish faster than orchestration | No run ledger, no bookkeeper, normal implementation and verification |
| Thin Workflow | Multi-step work benefits from visible phases, but durable artifacts would slow it down | update_plan, main-thread progress table, optional sidecar workers, no .workflow/ unless the run grows |
| Durable Workflow | User asked for ultracode/workflow, or the task needs resumability, auditability, multi-agent reduction, or cross-checking | Create .workflow/ultracode/<run-id>/, phase ledger, packets/results, progress dashboard, final report |
| Runner Workflow | A trusted workflow runner exists and the task is large, repeatable, or long-running | Use the runner after approval; mirror its state into the Ultracode ledger |
Escalate into Durable Workflow when any of these are true:
ultracode, workflow, dynamic workflow, subagent swarm, or broad fan-outStay Direct or Thin when:
Before a Durable or Runner workflow starts, show a compact run card and ask only for the meaningful approval. Do not ask again for routine worker events inside the approved envelope.
Ultracode run: react-to-solid-migration
Goal: Non-destructive React to Solid.js port of Excalidraw into solid-migration/
Mode: Durable Workflow
Budget: 4 concurrent agents, 16 total agents, 45 min first pass
Writes: solid-migration/** only
Network: off
Verification: typecheck, tests, migration report
| Phase | Purpose | Agents | Output |
| --- | --- | ---: | --- |
| 1 Inventory | Map files and dependencies | 4 | source map |
| 2 Patterns | Identify migration patterns | 4 | pattern guide |
| 3 Infrastructure | Create Solid shell and tooling | 3 | buildable app |
| 4 Core Port | Port shared logic | 3 | changed files + tests |
| 5 App Port | Port UI flows | 1 | changed files + tests |
| 6 Verify | Run checks and review | 1 | final report |
Ask before starting if the workflow exceeds the default budget, uses credentials, writes outside the repo, adds dependencies, starts long-running resources, or uses network-heavy research.
For Dynamic Workflow mode, keep a visible status surface in the main thread. Use a Markdown table in normal chat; use a compact text table if Markdown would wrap badly.
Recommended main-thread update:
Ultracode: react-to-solid-migration
10/16 agents complete | elapsed 5m30s | status yellow: infra fix in progress
| Phase | State | Agents | Elapsed | Evidence | Next |
| --- | --- | ---: | ---: | --- | --- |
| Inventory | done | 4/4 | 1m20s | results/001-* | accepted |
| Pattern Analysis | done | 4/4 | 2m45s | results/002-* | accepted |
| Infrastructure | running | 2/3 | 5m30s | 2 files changed | finish tooling |
| Migrate Core | pending | 0/3 | - | - | after infra |
Agent drill-down, when useful:
| Agent | Phase | State | Scope | Tools | Result |
| --- | --- | --- | --- | ---: | --- |
| infra:package.json | Infrastructure | done | package.json | 9 | scripts updated |
| infra:vite.config.ts | Infrastructure | running | vite.config.ts | 8 | checking aliases |
| infra:setupTests.ts | Infrastructure | blocked | test setup | 5 | needs jsdom choice |
Rules:
unknown instead of inventing tokens or tool counts.Preflight and capability detection
multi_agent_v1.spawn_agent, wait_agent, send_input, and close_agent are available before promising agent orchestration.tool_search only when the tools are not already visible.Draft the phase graph
Create the run ledger for Durable/Runner
.workflow/ultracode/<run-id>/ with workflow.md, state.json, journal.md, packets/, results/, integration.md, and final-report.md..workflow/ to local Git exclude, not tracked .gitignore, unless the user explicitly wants the ignore rule committed.Get approval once
Run phases dynamically
multi_tool_use.parallel for independent file reads and safe local commands.exec_command sessions for long-running commands only when the user approved the resource.state.json and journal.md at phase boundaries, blockers, integrations, verification results, and resource changes. Avoid per-tool-call bookkeeping.Reduce and cross-check
integration.md.Integrate and verify
final-report.md and run the artifact validator when a ledger exists.Defaults for Codex:
Claude Code may support much larger workflow runs, but Codex should start smaller because the parent still owns coordination and user-facing output.
Use artifacts because they preserve state, not because the skill says "workflow."
| Level | Files | Use when |
|---|---|---|
| None | none | Direct mode |
| Thin | update_plan only | visible progress is enough |
| Durable | workflow.md, state.json, journal.md, packets/, results/, integration.md, final-report.md | resumability, audit trail, multi-agent reduction |
| Runner Mirror | Durable files plus runner script/path/snapshot metadata | external runner or native workflow exists |
In Durable mode, update artifacts at meaningful lifecycle points only:
Avoid bookkeeping every command, file read, or intermediate thought.
When resuming an Ultracode run:
.workflow/ultracode/*/state.json.workflow.md, state.json, journal.md, integration.md, latest results/, and resource ledger.git status, relevant process/session checks, live browser/server/container state when applicable.A bookkeeper is optional in Codex and only useful when it is a real independent worker that can maintain Durable artifacts while the parent continues execution.
Use a bookkeeper when:
.workflow/ultracode/<run-id>/**Do not use a bookkeeper when:
If no bookkeeper is used, the parent owns compact artifact updates at phase boundaries. Workers never edit .workflow/ unless the packet explicitly assigns bookkeeper responsibility.
Every accepted result should include at least one of:
Do not accept:
Stop and ask when:
Read references/subagent-patterns.md for packet prompts and role details.
multi_agent_v1: use spawn_agent for bounded sidecars after user opt-in; use wait_agent sparingly; use close_agent after review; keep critical path local./workflows when the user wants that surface; mirror durable run facts only if cross-host portability matters.references/workflow-contract.md - durable ledger schema, state file, dashboard templates, packet/result templates, and final report format.references/subagent-patterns.md - Codex subagent prompts, dynamic phase patterns, bookkeeper rules, and conflict handling.references/risk-and-cost-gates.md - approval gates, budgets, safety constraints, resource cleanup, and verification standards.scripts/new-ultracode-run.mjs - optional helper that creates .workflow/ultracode/<run-id>/.scripts/verify-ultracode-run.mjs - validates expected artifacts before handoff.Do not use git add .workflow, git add -f .workflow, or include workflow artifacts in commits or PRs.