Artifact room
Skill kaskilling/agent-coordination-skills/skills/artifact-room
Focused agent-team coordination skills for deliberation, arbitration, parallel proof, diagnosis, handoffs, integration, and adversarial review.
npx -y skills add kaskilling/agent-coordination-skills --skill artifact-roomAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 29 days oldThe repository was created 29 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Coordinate multiple agents through one governed shared document, plan, decision record, or evidence board. Use when durable shared state should accumulate contributions across agents, provenance and conflicts must remain visible, and the result needs one coherent canonical artifact instead of separate reports.
SKILL.md
4.4 KB, 775 tokens by cl100k_base, as published. Nobody here has run it
Artifact Room
Build one durable artifact without concurrent-edit damage. Route every canonical change through exactly one writer.
Rules
- Assign one lead to own scope, the write protocol, conflict resolution, and the final answer.
- Preserve the user's mutation boundary. Do not turn review, diagnosis, or planning authority into permission to edit source files or external systems.
- Choose exactly one write protocol before contributors start:
- Canonical writer: one agent holds the write lease and all other agents send findings or proposed patches to that writer.
- Fragment and integrator: contributors write separate fragment files and exactly one integrator merges them into the canonical artifact.
- Never let two agents edit the canonical artifact concurrently, even when they own different sections.
- Give every contributor a distinct section, evidence obligation, and stop condition. Preserve source references, uncertainty, and open questions.
- Keep disagreements explicit until the lead resolves them. Do not overwrite or silently blend conflicting claims.
Workflow
1. Define the room
State the artifact's purpose and intended consumers in one sentence. Record:
- canonical artifact path or location
- required sections and acceptance criteria
- contributor ownership and mutation authority
- evidence and citation format
- unresolved-question and conflict format
- chosen write protocol and current writer or integrator
Create a small skeleton before delegation so contributors target a stable structure.
2. Collect contributions safely
Require contributors to read the latest canonical artifact before preparing their input. Ask each contributor to return:
- proposed content or fragment path
- supporting evidence and provenance
- assumptions and confidence
- conflicts with existing content
- open questions and validation not run
Under the canonical-writer protocol, contributors must not edit the artifact; send their proposals to the leased writer. Under the fragment protocol, assign non-overlapping fragment paths and forbid contributors from editing the canonical artifact or another agent's fragment.
Schedule work in waves when contributor count exceeds available capacity. While contributors work, inspect only unassigned seams or prepare consistency checks.
3. Integrate and resolve
Have the sole writer or integrator:
- verify each contribution against its evidence obligation
- merge compatible content without losing provenance
- record material conflicts instead of choosing silently
- ask the relevant owners one bounded clarification when needed
- let the lead resolve remaining conflicts against the acceptance criteria
After two exchanges that add no evidence, stop the dispute and record the missing proof or lead decision.
4. Validate and seal
Run one consistency pass over the complete artifact for:
- missing required sections
- terminology and decision consistency
- duplicate or contradictory statements
- broken references and lost provenance
- unresolved questions presented as facts
Ask intended consumers to read back the decisions, next actions, or operating steps they would take from the artifact. Fix material interpretation gaps through the sole writer, then mark the artifact sealed with remaining risks.
Degraded Path
If subagents or safe shared writes are unavailable, preserve the protocol:
- collect contributions serially while keeping earlier authorship and evidence
- use separate fragment files when possible
- otherwise keep one lead as the only canonical writer
State that the topology is degraded. Never compensate by allowing concurrent canonical edits.
Stop Conditions
Stop when all required sections meet their criteria, material conflicts are resolved or explicitly deferred, consumers can correctly use the artifact, and the sole writer or integrator seals it.
Stop earlier with a classified blocker when required evidence, write safety, or consumer readback is unavailable. Do not use this workflow for hot-line code collaboration, blind independent judgment, or isolated vertical implementation slices.
What ships with it: 1 file
220 B alongside SKILL.md
agents/
- openai.yaml220 B