| name | efficient-delegation |
| description | Use when running an expensive frontier model (Fable, Opus, or any future top tier) on codebase-heavy or token-heavy work — orchestrate with the frontier model and delegate research, coding, and testing to cheaper executors (Codex, Sonnet, Haiku). Triggers on "delegate this", "fan out subagents", "use cheap workers", "save frontier tokens", or any long multi-file task while the session runs a top-tier model. |
Spend frontier tokens on judgment only. Orchestrator = the session frontier model, whatever it is today (Fable now, anything tomorrow). Volume work → cheaper executors from the routing table.
<executor_table>
Edit this table to customize routing; rest of skill never changes.
| Work type | Executor | Notes |
|---|
| Guided code execution, plan exists | codex | Codex CLI, separate quota — spend freely |
| Bounded code edits needing Claude tools | sonnet | sub-agent model: sonnet |
| Research scans, repo inventory, docs extraction, log reduction | sonnet | parallel when slices independent |
| Mechanical: reformat, extract, rename, API glue | haiku | |
| Decomposition, architecture, conflict resolution, final review | session frontier model | never delegated |
Codex not installed → codex rows fall back to sonnet.
</executor_table>
1. Identify frontier-only decisions: decomposition, tradeoffs, ambiguity, risk, synthesis, final review. Keep inline.
2. Identify delegable slices: scans, inventory, docs extraction, test/browser passes, log reduction, bounded edits. Route per table.
3. Spawn parallel sub-agents for independent slices. Coupled or blocking work stays inline.
4. Require compact returns: findings, files + line refs, commands run, diffs, residual risk, stop conditions hit, anything the frontier model must decide.
5. Integrate, vet, review centrally before presenting.
<handoff_packet>
Every delegated prompt is self-contained — receiver has zero chat context:
- Objective + repo path
- In-scope files/surfaces + explicit out-of-scope
- Expected return format: files, line refs, commands, diffs, uncertainty
- Verification commands + what success looks like
- Stop conditions → stop and report, never improvise:
- live code contradicts the prompt
- a verification command fails twice after a reasonable retry
- work needs out-of-scope files
- cannot produce concrete evidence for a claim
</handoff_packet>
Sub-agent reports = leads, not facts. Before any high-impact use (PR, "done" claim, decision): reopen cited files, confirm line refs and failures, review the final diff against the task. Disagreeing agents → resolve at the frontier layer, never forward a guess.
- Never delegate the immediate blocker your next step depends on
- Never let two agents edit the same files concurrently
- Tiny task, or validation needing delicate judgment → keep inline
- RIFF projects: protocols/MODEL.md + /riff:wave already encode this routing — follow RIFF, do not double-orchestrate
<success_criteria>
- Frontier model produced no volume output (scans, logs, mechanical edits)
- Every delegated prompt carried the full handoff packet
- Every high-impact claim verified against actual files before reporting
</success_criteria>
