agentleFS
MesteriisMesteriis/Engineering-Bible-AI/skills/specialist-dispatch/SKILL.md

specialist-dispatch

Skill0 starsChanged 4 months agoMITView on GitHub

Selects scoped architecture, backend, frontend, security, data, DevOps, QA, docs, performance, or product roles.

We split this skill into the rules it gives an agent

9 rules. About 0 are common, 4 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this skill when a task benefits from a specialist lens but does not require installing a large external agent marketplace.

    line 16
  2. Distinctive:

    Do not spawn a specialist without a concrete question.

    line 40
  3. Distinctive:

    Prefer read-only specialist review unless implementation ownership is clear.

    line 42
  4. Very specific:

    Use agent-squad when multiple specialists will work in parallel.

    line 43
  5. Very specific:

    Use subagent-result-merge to consolidate outputs.

    line 44
  6. Very specific:

    When /multimodel is active, use budget models for patch proposals and stronger models for independent review.

    line 62
  7. Distinctive:

    Use a native custom subagent only when the installed host documents and functionally verifies a separate provider/model and its effective …

    line 68
  8. Very specific:

    Use agent-squad for concurrency limits and context-pack for exported data.

    line 78
  9. Very specific:

    Start with synthetic data, explicit iteration/output/request limits and an outer process deadline.

    line 86

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Specialist Dispatch

For parallel author workflows, select superpowers.dispatching-parallel-agents or superpowers.subagent-driven-development through be skills route ID --json and read the complete exposed provider. This skill retains Bible's specialist selection and worker authorization policy separately. Missing providers or delegation capabilities stay unresolved until the selected profile supplies them and host exposure is verified.

Use this skill when a task benefits from a specialist lens but does not require installing a large external agent marketplace.

This skill turns broad agent catalogs into a small Codex-native dispatch layer. It can be used with subagents when available, or as a checklist for the main agent when no subagent tool exists.

Specialist Profiles

Choose the smallest set:

  • architect: boundaries, dependency direction, module shape, migration seams;
  • backend: APIs, domain logic, persistence, concurrency, errors;
  • frontend: UI behavior, state, accessibility, visual quality;
  • security: auth, authz, secrets, parser risk, supply chain;
  • data: schemas, migrations, analytics, integrity, retention;
  • DevOps: CI, deploy, runtime config, observability;
  • QA: test plan, regression coverage, edge cases;
  • docs: user-facing and operator-facing documentation;
  • performance: evidence-led bottleneck analysis;
  • product: user workflow, scope, acceptance criteria.

Dispatch Rules

  • Do not spawn a specialist without a concrete question.
  • Give each specialist bounded files and expected output.
  • Prefer read-only specialist review unless implementation ownership is clear.
  • Use agent-squad when multiple specialists will work in parallel.
  • Use subagent-result-merge to consolidate outputs.

External Worker Roles

Map portable roles to verified providers and models in machine-local configuration. Default: keep the current selected model; multimodel and quorum are off. Preserve the parent's model, provider, and authorization. Only the user's /multimodel command selects the budget-author/strong-review workflow. /quorum selects its fixed reviewer roster and voting while keeping the author's model unchanged. Same-model specialist lanes activate neither mode. The following roles do not themselves authorize model changes:

  • scout: inspect bounded context and return a map of facts;
  • reviewer: review an allowed diff and related files with evidence;
  • tests-docs: propose tests or documentation as an artifact, without applying it;
  • patch-proposer: return a bounded code patch for the parent to inspect and apply;
  • local-helper: perform a bounded task in the locally verified inference mode.

When /multimodel is active, use budget models for patch proposals and stronger models for independent review. Role names do not prove availability, price or quality. Record actual routes and usage, exclude the proposer and integrator from approval votes, and never silently upgrade the implementation route. See Cross-Provider Review.

Use a native custom subagent only when the installed host documents and functionally verifies a separate provider/model and its effective permissions. Otherwise use an explicit, locally verified external CLI capability. A config entry is not evidence of a route. No hidden fallback: an unavailable requested model returns blocked or failed; a replacement must be explicitly selected and recorded. An unverified tool-calling model is only a text helper.

The main agent owns decisions and applies changes. External workers may not edit source, run arbitrary shell, merge, commit, push, install dependencies, or spawn other workers. Enforce limits in the runtime, not only in role text. Use agent-squad for concurrency limits and context-pack for exported data.

Alternative Runtime Pilots

Swarms or another programmatic runtime may be evaluated as an alternative external worker for fixed roles, not a provider switch inside the native parent's subagent tool. Export only the required operations and keep the existing subagent-result-merge contract, source scope and parent ownership. Start with synthetic data, explicit iteration/output/request limits and an outer process deadline; an async timeout alone may leave work running.

A function exported over MCP proves that function path, not model delegation. Separate server initialization, tool discovery, function invocation, actual model execution and parent-to-worker execution in the result. No automatic loops, global registration, new router or hidden provider fallback. Keep library import side effects, credentials and telemetry inside reviewed private runtime configuration. Do not stack frameworks without a measured benefit.

Specialist Prompt Shape

Specialist:
Question:
Scope:
Allowed files:
Forbidden files:
Evidence required:
Expected output:
Validation:
Task ID and base commit:
Snapshot state and hash:
Requested route and required runtime evidence:
Allowed tools and enforced permissions:
Timeout and retry budget:
Result: subagent-result-merge contract

Output

Report:

  • selected specialists;
  • why each was needed;
  • scope boundaries;
  • result or next routing skill.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.