agentleFS
S3n0001S3n0001/agent-skills/route-ultracode-subagents/SKILL.md

route-ultracode-subagents

Skill1 starsChanged 2 months agoMITView on GitHub

Claude Code only. Route Ultracode or dynamic-workflow stages to Sonnet 5 (easy), Opus 5 (medium), or Fable 5.1 (very complex) with exact worker prompts and resolved-model checks. Not ordinary single-agent work.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 4 are distinctive, and 1 is 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.

  • 1

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Check workflow and model availability when visible.

    line 14
  2. Distinctive:

    Read the dated model-routing evidence when availability, compatibility, privacy, or tier rationale matters, and refresh drift-prone facts …

    line 18
  3. Distinctive:

    Add an independent verifier when a medium or very complex implementation has a credible missed-failure path, and classify the verification …

    line 37
  4. Distinctive:

    When routed work spans sessions or owners, hand mission tracking to orchestrate-long-running-work.

    line 38
  5. Distinctive:

    Never delegate with "make the small fix".

    line 42

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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

Route Ultracode Subagents

Choose the lowest model tier that can reliably own each independently verifiable work item. Apply this only after Ultracode or a dynamic workflow is selected; it does not authorize settings changes, extra fan-out, or actions outside the user's authority.

Establish runtime truth

  1. Require an explicit human ultracode request, a direct dynamic-workflow request, or known /effort ultracode state. Difficulty alone never activates this skill.
  2. Request a full model ID on every routed stage: claude-sonnet-5, claude-opus-5, or claude-fable-5-1. If a newer ID in the same family is the documented current release, use it and say so.
  3. Check workflow and model availability when visible. Do not change CLAUDE_CODE_SUBAGENT_MODEL, an allowlist, provider settings, or the session model without authorization.
  4. Distinguish requested from resolved models. Environment overrides and allowlist substitution can defeat routing; inspect runtime evidence or mark the resolved model unverified.
  5. Apply confidentiality, retention, safety, provider, and budget constraints before capability routing.

Read the dated model-routing evidence when availability, compatibility, privacy, or tier rationale matters, and refresh drift-prone facts against official documentation.

Classify each work item

Classify the smallest coherent unit one worker can own and verify. Classify mixed requests stage by stage. Use the highest applicable risk or reasoning signal; line count is only a hint.

ClassModelEvidenceExample
Easyclaude-sonnet-5One known seam, precise behavior, reversible, usually one or two files, no consequential contract, security, data, migration, or concurrency behaviorChange two known strings and run the focused check
Mediumclaude-opus-5A bounded feature, fix, review, or refactor with non-trivial reasoning across known components and understood failure and acceptance surfacesAdd validation across a handler, service, and tests
Very complexclaude-fable-5-1Interacting systems, uncertain ownership, long-horizon work, large migration, difficult architecture, or consequential security, data-integrity, concurrency, compatibility, or rollback behaviorImplement a staged service split with mixed-version proof

Escalate small but consequential edits: two authorization or migration lines are not easy. Classify discovery, implementation, and verification separately; a cheap discovery stage must not choose the implementation model.

Shape the workflow

  • One worker per coherent outcome, not per file. Use the minimum useful agent count.
  • Parallelize only independent inputs with disjoint write sets. Serialize dependencies and shared-file edits.
  • Give every worker an outcome, scope, preserve constraints, verification duty, and stop condition. Inspect its result before integration.
  • Add an independent verifier when a medium or very complex implementation has a credible missed-failure path, and classify the verification on its own difficulty.
  • When routed work spans sessions or owners, hand mission tracking to orchestrate-long-running-work; this skill owns only per-stage routing.

Give Sonnet exact instructions

Never delegate with "make the small fix". For every Sonnet item, fill this contract with concrete evidence:

Outcome: <one observable result>
Scope: <exact paths, symbols, or permitted write set>
Required change: <precise before/after behavior>
Preserve: <behavior, interfaces, formatting, and unrelated work that must not change>
Evidence: <focused command or inspection and the passing signal>
Return: <changed files, concise result, exact checks and outcomes, remaining uncertainty>
Stop if: <scope expands, evidence conflicts, or a consequential choice is missing>

Give Opus the same outcome, boundaries, integration contract, and falsifiable verification with method autonomy. Give Fable the full mission contract: architecture evidence, invariants, dependencies, failure and rollback semantics, integration ownership, and end-to-end acceptance. Verbosity is not evidence.

Stop and ask when

  • routing would require changing a model setting, allowlist, or provider configuration;
  • a medium or very complex stage can only resolve to a weaker model and the user has not accepted that;
  • the request is not actually an Ultracode or dynamic-workflow request.

Done when

The integrated outcome is proven by inspecting the combined diff or artifact and running focused checks; every subagent returning is not completion. Report each item's class, requested and resolved model, result, verification, and any substitution, refusal, fallback, or unresolved risk. Never claim a requested model ran without evidence.

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.