delegate
Multi-agent delegation policy for Grok Build: when DeepSeek (main) should spawn specialized subagents, which model/type to use, and when to run them in…
Skill by TheNonFungibleHuman
Chalk Labs
Need help deploying delegate?
Chalk Labs sets up agents like this inside real teams: connected to your tools, tested on your data, with a human in the loop where it matters.
Install
Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.
npx agentmarket install skill/thenonfungiblehuman-delegate
Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: research, agent-building
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
Permissions
What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.
- Tools
- No tools found
- Secrets
- None asked for
- Install scripts
- None found
Network hosts it talks to 1
codeload.github.com· found in install command
Files it touches
~/.claude/skills/delegate
What's inside
- Skill name
delegate- Description
- Multi-agent delegation policy for Grok Build: when DeepSeek (main) should spawn specialized subagents, which model/type to use, and when to run them in parallel. Use when starting multi-step work, Paper + vision, explore + implement, review before merge, or when the user says /delegate, "use subagents", "run in parallel", or "multi-agent".
- License
- MIT
- Instructions
- 257 lines in SKILL.md · view on GitHub
Files (1)
- SKILL.md
SKILL.md (first lines)
# Delegate — main DeepSeek + specialist subagents You are the **foreman** (usually **DeepSeek Flash** on the main session). Specialists are **subagents** with their own context. You integrate **short summaries** only. This is **policy you follow**, not a hidden auto-router. Spawn when the table says so; do not fan out for trivial work. ## Parallelism — yes, Grok CLI supports it | Capability | Supported? | |------------|------------| | Multiple subagents at once | **Yes** | | How | Issue **multiple** `spawn_subagent` calls in one turn with `background: true` (default), then wait / collect results | | Nested subagents (child spawns child) | **No** — max depth 1 | | Same files edited by two writers in parallel | **Avoid** — conflicts / thrash | **Parallel when:** read-only, different modalities, or independent scopes (e.g. explore auth + explore billing; vision QA + web research). **Sequential when:** one depends on the other, or both write the same paths / same Paper artboard. ```text # Good parallel fan-out (one parent turn) spawn_subagent(explore, "Find auth middleware…") # background spawn_subagent(explore, "Find rate-limit config…") # background → await summaries → main DeepSeek implements (and reads screenshots itself) ``` ## Seats (your harness) | Role | Config / model id | Notes | |------|-------------------|--------| | Main bulk | `deepseek-flash` | Bulk foreman / implementer | | Bulk + think (default) | `deepseek-flash-high` | Same model, higher effort. The config default daily driver. | | **Judgment / hard** | **`deepseek-pro`** | Planning, nasty bugs, review. Also the `plan` subagent and both personas. Text-only. | | Overflow bulk | `glm-flash` | Z.ai GLM 5.3 Flash (PAYG), 1M context, low effort. Alternative bulk seat; same rules — the pro seat still plans and debugs. | | Explore children | `deepseek-flash` via `[subagents.models]` | Read-only search | | **Plan children** | **`deepseek-pro`** via `[subagents.models].plan` | **Always the pro seat for real planning** | | Vision | `deepseek-flash` or `glm-flash` on the current seat | Screenshots, Paper, mockups. Pro is text-only. **No Gemini, no xAI seats.** | | **Nasty bugs / hard debug** | **`deepseek-pro`** | **Must spawn the pro seat** — not Flash alone | | Review / architecture | `deepseek-pro` | Security, design judgment, go/no-go | | Web facts | native `web_search` / `web_fetch` (pinned to `deepseek-flash`) | Exa MCP for deep research or fallback | | Images create/edit | `image_gen` / `image_edit` tools | Not a chat subagent | **No local proxies.** 8787, 8788, 8789 and 8791 are all retired — never start one to work around a seat error. ### Specialty split (explicit) | Specialty | Model | When | |-----------|--------|------| | **Volume coding** | DeepSeek Flash | Default implement, tests, refactors | | **Planning** | **DeepSeek Pro** | User wants a plan, multi-step design, “how should we build X”, architecture before code | | **Nasty / hard bugs** | **DeepSeek Pro** | Heisenbugs, flaky, multi-system, security, “can’t repro”, Flash already failed, race/concurrency, production-only | | **Vision** | `deepseek-flash` or `glm-flash` | Pixels / screenshots. Pro is text-only. Gemini unplugged. | | **Explore map** | DeepSeek explore | “Where is X?” only | Spawnable types (probe-confirmed on this build): `explore`, `general-purpose`, `plan`. The `[subagents.personas.*]` blocks in config are prompt overlays only — they do **not** create spawn types, so never pass `subagent_type: reviewer` or `planner` (the call fails with "Unknown subagent type"). Pin **`model`** on spawn (e.g. `model: deepseek-pro`) and put specialty rules in the **child prompt**. For `plan`, config already sets the model to DeepSeek Pro. ## Task → specialist table | User / situation | Do on main (DeepSeek)? | Spawn | Type | Model | Parallel OK? | |------------------|------------------------|-------|------|-------|--------------| | Trivial edit, rename, one-liner | **Yes alone** | — | — | — | — | | Multi-file implement, tests | **Yes** (or drafter child) | optional drafter | `general-purpose` | `deepseek-flash` | No if same files as main | | “Where is X?” / map codebase | Light grep OK | **explore** | `explore` | deepseek | **Yes** × independent queries | | **Any non-trivial plan / design / “how should we…”** | **No — do not plan alone on Flash** | **planner** | `plan` or `general-purpose` | **`deepseek-pro`** | Usually 1 (optionally parallel with explore) | | **Nasty / hard / subtle bug** | **No — do not debug alone after one weak try** | **debugger / reviewer** | `general-purpose` | **`deepseek-pro`** | 1 (explore may run first or in parallel) | | Screenshot / Paper visual QA | **Yes — read on `deepseek-flash` / `glm-flash`** | — | — | — | n/a | | Bug screenshot → fix | **Yes** (read image on main) | pro seat if nasty | `general-purpose` | **`deepseek-pro`** | Sequential for nasty | | Paper design loop | Paper MCP on main; DeepSeek reads screenshots | — | — | — | Not two Paper writers | | Security / merge review | No | **reviewer** | `general-purpose` | **`deepseek-pro`** | After implement | | Architecture decision | No | **planner** | `plan` / `general-purpose` | **`deepseek-pro`** | 1 | | Fresh API/docs facts | Call tool | — | — | `web_search` | N/A | | Image generation | Tools on main | — | — | Imagine tools | Parallel gens OK | ## Child prompt contract (always) Every spawn prompt must include: 1. **Goal** (one sentence) 2. **Scope** (paths, artboard, constraints) 3. **Output format** (bullets / files changed / go-no-go) 4. **Do not** re-dump the whole repo; return a **short summary** for the parent ### Vision child (template) Do **not** spawn a Gemini vision child. Gemini seats are unplugged. Read images on `deepseek-flash` or `glm-flash` (the pro seat is text-only). ### Reviewer child (template) ```text Strict review. Check correctness, edge cases, security, match to goal. Return: summary, blockers, nits, go/no-go. Cite paths. No fluff. ``` `model`: **`deepseek-pro`** · high effort if available. ### Nasty-bug / hard-debug child (template) — **always DeepSeek Pro** ```text You are the senior debugger (DeepSeek V4.1 Pro) on a hard debug. Parent is Flash (implementer only). Symptoms: <paste> What we tried: <paste> Reproduce: <commands/paths if known> Return ONLY: - Root-cause hypothesis (ranked, with confidence) - Evidence to gather / commands to run - Minimal fix plan (files + steps) - Risks / regressions Do not implement large rewrites; hand a clear fix plan back to the parent. ``` `model`: **`deepseek-pro`** · `subagent_type`: `general-purpose` · high effort if available. **Triggers (spawn the pro seat — do not keep thrashing Flash alone):** - User says nasty / hard / flaky / heisenbug / race / production-only / security bug - Same issue failed **twice** on Flash - A Flash fix broke a p2p test (the measured conversion case) - Multi-service or concurrency / memory corruption style symptoms - You are guessing without a clear root cause
From TheNonFungibleHuman/multi-model-agent-harness, MIT.
More from TheNonFungibleHuman/multi-model-agent-harness
- 0.565score
handoff
TheNonFungibleHuman
Write a compact handoff document so another Grok session (or agent) can continue this work.
SkillAFreshness: AClaude Code +9 - 0.565score
route
TheNonFungibleHuman
Multi-model routing policy for this harness.
SkillAFreshness: AClaude Code +9 - 0.565score
capture-lesson
TheNonFungibleHuman
Capture a durable lesson from a bug, failed approach, or correction so future sessions improve.
SkillAFreshness: AClaude Code +9 - 0.565score
product-build
TheNonFungibleHuman
End-to-end product build conductor.
SkillAFreshness: AClaude Code +9 - 0.565score
pull-handoff
TheNonFungibleHuman
Load the latest (or named) handoff brief from ~/.grok/handoffs into this session so work can continue.
SkillAFreshness: AClaude Code +9 - 0.558score
profile-full
TheNonFungibleHuman
Switch Grok to the full config profile (design MCP on, broader skill catalog).
SkillAFreshness: AClaude Code +9 - 0.558score
profile-lean
TheNonFungibleHuman
Switch Grok to the lean config profile (fewer skills, Neon off, lower base tokens).
SkillAFreshness: AClaude Code +9 - 0.556score
shelf-off
TheNonFungibleHuman
Disable all optional skill shelf packs (design, Matt, video, extras, firecrawl), leaving the core pack on.
SkillAFreshness: AClaude Code +9 - 0.553score
shelf-status
TheNonFungibleHuman
Show whether core / design / Matt / video / extras / firecrawl skill shelf packs are enabled.
SkillAFreshness: AClaude Code +9 - 0.553score
mcp-design-off
TheNonFungibleHuman
Disable Paper + Open Design MCP and design skills for a leaner token baseline.
SkillAFreshness: AClaude Code +9 - 0.553score
shelf-pstack-on
TheNonFungibleHuman
Enable the pstack (poteto-mode) skill pack from the shelf.
SkillAFreshness: AClaude Code +9 - 0.553score
profile-status
TheNonFungibleHuman
Show whether the lean or full Grok config profile is active.
SkillAFreshness: AClaude Code +9
Signals
- Stars
- 0
- Forks
- 0
- Open issues
- 0
- Last commit (this folder)
- Oct 19, 2026 (12 days ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- MIT
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from TheNonFungibleHuman
- 0.565score
route
TheNonFungibleHuman
Multi-model routing policy for this harness.
SkillAFreshness: AClaude Code +9 - 0.565score
handoff
TheNonFungibleHuman
Write a compact handoff document so another Grok session (or agent) can continue this work.
SkillAFreshness: AClaude Code +9 - 0.565score
capture-lesson
TheNonFungibleHuman
Capture a durable lesson from a bug, failed approach, or correction so future sessions improve.
SkillAFreshness: AClaude Code +9 - 0.565score
pull-handoff
TheNonFungibleHuman
Load the latest (or named) handoff brief from ~/.grok/handoffs into this session so work can continue.
SkillAFreshness: AClaude Code +9 - 0.565score
product-build
TheNonFungibleHuman
End-to-end product build conductor.
SkillAFreshness: AClaude Code +9 - 0.558score
profile-full
TheNonFungibleHuman
Switch Grok to the full config profile (design MCP on, broader skill catalog).
SkillAFreshness: AClaude Code +9
Done for you
Rather not set it up yourself?
Chalk Labs scopes, builds and runs agents for your team: the task audit, the eval harness, the integrations and the weekly ops report. Fixed scope, priced before we start.
See how we build AI agents →Read the case studies
Chalk Labs is a two-founder AI and growth studio. We build AI agents, CRMs and MVPs for Web3 and AI companies, and have shipped 50+ product and token launches.
- launches shipped
- 50+
- launches shipped
- media placements
- 500+
- media placements
- founders on every account
- 2
- founders on every account
Press we've placed: Forbes · CoinDesk · Decrypt · Cointelegraph · The Block
Rahil Jain, Co-founder, Chalk Labs. Ex-Goldman Sachs engineer. You talk to the person who builds it.
Get delegate running in your business
Tell us what you want to automate. A founder replies within one working day with a plan, not a pitch deck.
Cite this page
delegate skill. Agent Mart. https://agentmart.dev/skills/thenonfungiblehuman-delegate (last updated 2026-10-24)