orchestrate · prathamdby — Claude Code skill: delegate work, verify it
A skill that runs a parent agent as coordinator: it scopes the job, hands chunks to subagents, then checks and merges what comes back.
npx skills add prathamdby/skills --skill orchestrate -a claude-codeWhat orchestrate does
Turn the skill on and the parent agent stops touching product files. It writes briefs instead, picks the cheapest model that can handle each chunk, and keeps a ledger line after every step: task | criteria | owner/model/strikes | artifacts/evidence | verified | blocked | waivers. The mode stays active until you say “stop orchestrating” or “normal mode”.
Verification is the point. Self-reports from delegates don't count, so the parent inspects diffs, runs the relevant tests and builds, and keeps evidence pointers for later. A failed check adds a strike to that owner: re-brief first, swap the delegate second, block third. Fixing a delegated chunk yourself is off the table.
Resuming works from diffs, test results and artifacts rather than the model's memory of the session. Delegation needs agent types you can actually reach and a model choice per chunk; with nobody to delegate to, the skill offers to leave the mode instead of continuing alone. disable-model-invocation: true in the frontmatter means you start it, not the model.
What it helps with
- split a large task into chunks with disjoint write boundaries
- verify subagent output against diffs, tests and builds
- integrate results from several agents into one deliverable
- track owners, strikes and blocked chunks in a ledger
- rebuild session state from artifacts after a pause
Example prompts for the agent
- Turn on orchestration mode and split this task into independent chunks
- Verify the subagent results against the diffs and tests, show me the ledger
- Report what is verified, blocked and waived
How to install orchestrate
1Skills CLI (Vercel)
npx skills add prathamdby/skills --skill orchestrate -a claude-code2Manually: put the files in place
~/.claude/skills/orchestrate/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/orchestrate/
1Skills CLI (Vercel)
npx skills add prathamdby/skills --skill orchestrate -a codex2Manually: put the files in place
.agents/skills/orchestrate/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add prathamdby/skills --skill orchestrate -a cursor2Manually: put the files in place
.cursor/skills/orchestrate/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install prathamdby/skills orchestrate2Manually: put the files in place
.github/skills/orchestrate/SKILL.md1Skills CLI (Vercel)
npx skills add prathamdby/skills --skill orchestrate -a gemini-cli2Platform CLI
gemini skills install https://github.com/prathamdby/skills --path skills/engineering/orchestrateManually: put the files in place
.devin/skills/orchestrate/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add prathamdby/skills --skill orchestrate -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about orchestrate
How do I install orchestrate in Claude Code?
Run npx skills add prathamdby/skills --skill orchestrate -a claude-code. The skill file lands at ~/.claude/skills/orchestrate/SKILL.md.
What does orchestrate do?
It puts the parent agent in charge of scoping, briefing and verifying delegated chunks, then integrating the results. The parent does not edit product files itself.
What happens when a delegate fails verification?
That owner takes a strike: re-brief first, replace second, block third. The skill forbids the parent from repairing the chunk on its own.
Source text
SKILL.md · MIT
--- name: orchestrate description: Coordinate delegated work with parent-owned scope and verification. disable-model-invocation: true --- # Orchestrate Parent scopes, briefs, verifies, integrates, and judges. Delegates research/edit; parent does not edit product files or delegate its own responsibilities. Active until "stop orchestrating" or "normal mode". Record after each step: `task | criteria | owner/model/strikes | artifacts/evidence | verified | blocked | waivers`. On resume rebuild from diffs, results, and tests, not memory. Revalidate evidence affected by changed files/tools, refresh roster, and reassign disappeared delegates. Defer unrelated work unless the user replaces or queues the active task. ## 1. Muster and scope Discover available agent types, write scope, and model selection. Pick the cheapest capable model per chunk and record it. Missing delegation blocks; offer to leave mode instead of proceeding solo. Turn each outcome into observable criteria. Correct reversible flaws with evidence and record requested/corrected scope; pause before irreversible or materially broader corrections. Done when roster and all criteria are recorded. ## 2. Brief and dispatch Partition independently verifiable, disjoint write boundaries. Each brief has goal, reason, context pointers, constraints, write scope, expected evidence, completion criterion. Delegates cannot delegate. Launch independent chunks in one wave, explicitly choosing model when supported; queue dependencies/overlaps. Prepare verification and inspect completed results. Done when each chunk has one owner and is running, finished, or gated. ## 3. Verify and integrate Self-reports are unverified. Inspect files/diffs, run relevant tests/builds, and check interactions. Preserve evidence pointers so resumed checks focus on affected artifacts; unchanged prose claims do not replace actual tests. Each failed check increments that owner's strikes: re-brief first, replace second, block third. Never fix a delegated chunk yourself. Integration failure gets a coordination brief naming affected owners/evidence; apply the same ladder to each implicated owner's running strike total. Done when every criterion and integration is VERIFIED, evidenced BLOCKED, or explicitly waived by the user by name in the ledger. ## 4. Report Lead with overall and integration result. List verified, blocked, and waived criteria with evidence. For corrected scope report "Deviation: requested X; evidence Y; delivered Z." Done when every claim has current-run evidence and no delegate is untracked.
© prathamdby · MIT · prathamdby/skills
Similar skills
All skills for Agent setup: memory, context, orchestration →orchestrate by other authors comes first.