---
name: resource-planning
description: Plan translation-agent resources using language, domain, tool and capacity evidence. Use when selecting specialists, allocating target locales, preserving translator/reviser independence or revising a team after a capability gap.
---

# Resource Planning

Read the [shared expert contract](../.shared/references/expert-contract.md).
Acknowledge this skill, actual actor identity, assigned scope, inputs and gaps.
Assess capability and propose assignments; do not create qualification claims.

## When to invoke

Use when a coordinator needs a feasible team, when a language pair or specialist
route is unproven, or when capacity constraints require sequencing. You can assess
available resources directly without a completed translation job.

## Inputs and evidence

- Approved products, source/target locales, domain, risk and review coverage.
- Actual host delegation tools, slot limits and any context/resource constraints.
- Available role skills, approved tool access and permitted information flow.
- Language-pair/domain pilot results, prior work samples and known limitations.
- Local runner readiness, model/context limits and measured inference capacity.

## Workflow

1. Decompose the brief into competence requirements. Separate language fidelity,
   target-language expression, subject knowledge, file/media production and
   accessibility. A single model label does not establish all of these abilities.
2. Inspect the actual host capabilities and available role skills. Map each
   proposed assignment to a real agent slot and role-skill path. Record unavailable
   delegation or media capabilities as gaps rather than inventing tools or IDs.
3. Compare competence evidence with the task. Prefer representative samples and
   explicit error analysis over unverified fluency claims. Propose a pilot for an
   unqualified pair or domain; identify what evidence would permit scaling up.
4. Design independence deliberately. Assign a different actual agent to bilingual
   revision of material drafted by a translator. Do not reuse that translator's
   identity under a new role label. Separate other functions when risk or scope
   requires it; combine compatible functions transparently for small jobs.
5. Schedule around dependencies and slots. Queue work instead of spawning a full
   notional agency at once. Parallelize independent analysis; keep shared runtime
   mutations and same-artifact edits serialized by the coordinator. Respect the
   runner's measured capacity rather than launching multiple model loads.
6. Return a proposed roster and gap plan. Ask the coordinator to dispatch actual
   agents with explicit skill loading. Reassess an assignment when the returned
   work or pilot reveals a competence, context or access limitation.

An agent that writes polished French may still omit source conditions. Judge its
assigned work against a bilingual sample and choose independent revision; do not
promote that fluency observation into a professional qualification record.

## Outputs and handoff

Use the [expert handoff](../.shared/templates/expert-handoff.md). Return a roster
with role/skill, locale/domain, proposed actor, required inputs/tools, competence
evidence, independence constraints, slot sequence, output ownership and gaps.

Distinguish proposed resources from actually dispatched agents. Once supplied by
the coordinator, record real agent IDs and any reassignment reasons.

## Completion evidence

The roster can execute within actual host and runner capacity. Required
independence is explicit, unsupported functions have a route or blocker, and
unqualified language pairs remain labeled until pilot evidence supports them.

## Must not

Do not claim human credentials, certify a model, invent agent availability or
promise privacy from a local runner alone. Do not silently replace a required
specialist with a generic pass or use the original translator as their own
independent reviser.