Sub-Agent Forced

A Claude Code skill that turns the main agent into a pure orchestrator: it never does real work itself, it only decomposes, dispatches to subagents, independently verifies their output, and reports.

The point isn't speed — it's a clean main context (the doing happens in subagents) and sharp judgment (you check results with your own hands instead of trusting reports).

What it does

Under this skill the main agent is split into verify vs produce:

✅ Main agent MAY do	❌ Main agent MUST delegate
One-line factual answers, clarifying questions	Writing or editing any code / file
Decompose & dispatch subagents	Designing architecture, schemas, APIs
Verification only: grep, build, test, read, diff	Brainstorming / spec / approach work
Synthesize subagent results into a report	Implementing a feature or fix
Decide what to delegate next	Writing tests

The four roles

1. Decompose — break the request into well-bounded units; independent → parallel, dependent → sequence.
2. Dispatch — one subagent per unit, each with a self-contained brief and a clear definition of done.
3. Verify (with your own hands) — confirm at least one load-bearing fact yourself. A green self-check, not a green self-report.
4. Report & synthesize — what was done, how it was verified, what's unconfirmed, what's next.

The default 5-stage pipeline

For non-trivial (multi-step, multi-file) requests, the skill automatically applies a 5-stage pipeline: Research → Classify → Organize → Plan → Execute.

Each stage is delegated to subagent(s), and the main orchestrator verifies one concrete fact between stages before proceeding. Trivial requests (single file, one-line answers) bypass the pipeline; complex requests benefit from structured decomposition and inter-stage verification.

• Research: Investigate context, current state, constraints, and precedents in parallel.
• Classify: Group findings into meaningful categories.
• Organize: For each category, structure the work details and dependencies.
• Plan: Generate actionable work units with explicit dependencies.
• Execute: Dispatch and run each unit; verify results per the verification rule.

See SKILL.md for detailed stage definitions, flow diagrams, and gate criteria.

The heart of it — the verification rule

A subagent saying "done, tests pass, 0 errors" is a claim, not a fact. Before reporting any unit as done, you independently confirm a load-bearing fact (run the build, run the tests, grep the change, read the file, diff the result). If your check disagrees with the report, the unit isn't done — re-dispatch with the discrepancy spelled out.

Install

Claude Code discovers skills under ~/.claude/skills/. Drop the folder in:

git clone https://github.com/jangfolk/sub-agent-forced.git
mkdir -p ~/.claude/skills/sub-agent-forced
cp sub-agent-forced/SKILL.md ~/.claude/skills/sub-agent-forced/SKILL.md

Then invoke it via the Skill tool (or /sub-agent-forced if your setup exposes it as a command). Once active, the main agent will delegate everything and only verify + report.

License

MIT