coord-split
Skill0 starsChanged 47 days agoApache-2.0View on GitHub
Decompose an oversized or badly-shaped coord task into N dependent children. Interviews the user via grill-me to find the natural fault lines, creates child tasks chained with depends_on, then archives the parent. Use when the user says "/coord-split", "this task is too big", or "split <id>".
- Commits and pushes
In this repository20 other files it gives its agents.
- coord-archive-indexSkill
- coord-assignSkill
- coord-auditSkill
- coord-checkSkill
- coord-dirtySkill
- coord-discardSkill
- coord-helpSkill
- coord-initSkill
- coord-loop-doctorSkill
- coord-promoteSkill
- coord-refineSkill
- coord-requeueSkill
- coord-reviewSkill
- coord-shape-reviewSkill
- coord-shapeSkill
- coord-statusSkill
- coord-tokensSkill
- coord-unblockerSkill
- merge-to-mainSkill
- rebase-to-mainSkill
The file
We split this skill into the rules it gives an agent
12 rules. About 0 are common, 2 are distinctive, and 10 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
10
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 29
Per-child agent and model: default to Codex unless the user says otherwise.
- Distinctive: line 31
Stop the interview when the children are crisply defined (typically 2–5 children).
- Very specific: line 39
Run coord new for each child with explicit --complexity, --kind, --reasoning_effort, --model_claude, --model_codex, optional --agents, …
- Very specific: line 45
If any coord new exits non-zero, stop immediately.
- Very specific: line 48
Run coord promote <child-id> for every child after coord new.
- Very specific: line 49
No coord code (worker.sh, cmd_promote, any launchd job) auto-flips shaping → pending when deps clear.
- Very specific: line 66
Never use --force or --no-verify, except the step 5 parent-close --force.
- Very specific: line 67
Never edit task files directly.
- Very specific: line 68
Do not add roles.architect when splitting normal work.
- Very specific: line 69
Do NOT use the umbrella pattern (parent kept alive as a tracker).
- Very specific: line 70
If coord new rejects a child for shape warnings, treat it as the user's signal that the child is itself too broad — re-grill that child …
- Very specific: line 71
Do not stage anything by hand.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Break a too-broad coord task into properly-shaped children. Companion to /coord-shape (which shapes one task) — /coord-split is for when shaping reveals the task should have been multiple tasks.
Project-root preflight:
- Resolve the canonical project checkout before reading, creating, updating, staging, committing, or pushing:
COORD_MAIN=$("${COORD_TOOLS:-$HOME/Projects/coord-wright}/bin/coord-project-root") && cd "$COORD_MAIN" - This is required when the skill is invoked from a sibling task worktree; split child tasks and parent updates belong to the main queue that the launchd worker reads.
Syntax
/coord-split <task-id>
Workflow
- Load the parent.
python3 "${COORD_TOOLS:-$HOME/Projects/coord-wright}/bin/coord" show <id>→ capture task body, current subtasks, scope, acceptance, depends_on, kind, complexity.- Refuse if status is
done,archived, or alreadyneeds-brainstorming— those don't need splitting.
- Interview the user via grill-me.
- Invoke the
grill-meskill with the parent's title, current scope, and current subtasks as context. - The interview must surface:
- Fault lines: where does the task naturally divide? Each child should have one main outcome and one final verification path.
- Order and dependencies: which child must finish before the next?
- Per-child agent and model: default to Codex unless the user says otherwise.
- Per-child complexity, kind, reasoning_effort: inherit from parent or downgrade if the child is genuinely smaller.
- Stop the interview when the children are crisply defined (typically 2–5 children).
- Invoke the
- Create each child task.
- For each child, write four files before running
coord new:/tmp/coord-split-<parent-id>-<n>-plan.txt— concrete Plan (lead with prose; use###for any sub-heading such as### Subtraction analysis, never##). Without itcoord newseeds a placeholder Plan and step 4promoteexits 3./tmp/coord-split-<parent-id>-<n>-subtasks.txt— subtask block/tmp/coord-split-<parent-id>-<n>-acceptance.txt— acceptance bullets/tmp/coord-split-<parent-id>-<n>-scope.txt— scope paths (if needed)
- Run
coord newfor each child with explicit--complexity,--kind,--reasoning_effort,--model_claude,--model_codex, optional--agents, optional--assigned, optional--rolesfor explicit review or an exceptional architect handoff,--depends_on=<previous-child-ids>to chain them, and:--set-plan=@/tmp/coord-split-<parent-id>-<n>-plan.txt--set-subtasks=@/tmp/coord-split-<parent-id>-<n>-subtasks.txt--acceptance=@/tmp/coord-split-<parent-id>-<n>-acceptance.txt--scope=@/tmp/coord-split-<parent-id>-<n>-scope.txt(omit if no scope needed) - Capture each new child id from coord's stdout, then set
coord update <child-id> --scope-budget-loc="<band>"and run thecoord-reviewskill on the child before promoting it. - If any
coord newexits non-zero, stop immediately. Do NOT archive the parent; report which child failed and why so the user can fix the inputs and retry.
- For each child, write four files before running
- Promote ALL children to pending, not just the chain head.
- Run
coord promote <child-id>for every child aftercoord new. The worker'sdependency_blockerscheck incmd_pickupalready enforces ordering by skipping tasks with unmet deps; leaving siblings inshapingstrands them until someone manually re-runs/coord-promote. - No coord code (worker.sh, cmd_promote, any launchd job) auto-flips
shaping → pendingwhen deps clear.
- Run
Archive the parent — recipe for
needs-brainstormingor incomplete-subtasks state.- A plain
--status=donefails fromshaping,pending, orneeds-brainstorming(no edge todone) and while subtasks are incomplete (they were migrated to children, not finished). Close the parent in one forced call, which also records the child ids:python3 "${COORD_TOOLS:-$HOME/Projects/coord-wright}/bin/coord" update <parent-id> --status=done --force --add-issues="Split into: <child-id-1>, <child-id-2>, ..." --forceis the right authority here because the parent's subtasks are genuinely incomplete (migrated to children). TheRulessection's general "no--force" applies to normal closes; this is the documented/coord-discardcarve-out. Do not route the parent throughpendingfirst: that makes it runnable (a live worker can pick it up) and runs the runnable-shape gate.coord update --status=donearchives the parent in the same call. Closing a never-startedshaping/needs-brainstormingparent skips the model/effort baseline check, so an off-baseline draft still closes.
coord new,coord update, andcoord promoteeach commit and push their own change. Do not run manualgit add,git commit, orgit push; if a push failed,git rev-list --count origin/main..mainshows the unpushed commits.- A plain
- Report.
split: <parent-id> -> <child-id-1>, <child-id-2>, ...- For each promoted child:
promoted: <child-id> - Final line: push outcome (any
git push failedoutput fromcoord).
Rules
- Never use
--forceor--no-verify, except the step 5 parent-close--force. - Never edit task files directly; always use
coord newandcoord update. - Do not add
roles.architectwhen splitting normal work. Use child tasks chained withdepends_onas the decomposition mechanism; reserve architect roles for explicit in-loop design handoffs. - Do NOT use the umbrella pattern (parent kept alive as a tracker). The children's
depends_onchain encodes ordering; an umbrella duplicates state. - If
coord newrejects a child for shape warnings, treat it as the user's signal that the child is itself too broad — re-grill that child rather than bypassing the warning. - Do not stage anything by hand;
coordcommits the parent and child files itself.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

