> ## Documentation Index
> Fetch the complete documentation index at: https://aidocs.ethanbragdon.icu/llms.txt
> Use this file to discover all available pages before exploring further.

# FORGE Limits & Best Practices

> Design FORGE jobs that fit workspace, runtime, concurrency, validation, artifact, and credit limits without sacrificing reliability.

# FORGE limits & best practices

FORGE jobs can use substantially more compute and storage than an ordinary chat request. Build Mode therefore operates within explicit platform, plan, organization, and workload limits.

## Common limit categories

* concurrent builds
* build runtime
* model/tool calls
* workspace size
* maximum file count
* per-file size
* repair attempts
* artifact size
* artifact retention
* credits or hard usage ceilings

The current product surfaces and Control Center configuration are the source of truth for exact values.

## Keep builds scoped

Prefer one coherent feature or project objective per build. Very large requests are harder to plan, validate, repair, review, and retry.

For a large migration, use staged builds with explicit checkpoints rather than asking FORGE to rewrite an entire product in one opaque run.

## Preserve source state

When modifying an existing repository, record the branch/revision used by the build. If the source changes significantly, start a fresh build against the updated state.

## Require validation appropriate to the risk

A UI-only change and a database migration do not need the same review strategy. Ask for the strongest validators the environment can actually run and perform target-environment checks afterward.

## Keep repair bounded

Repeated automated repair can consume credits while moving farther from the original architecture. If the same class of failure repeats, stop and inspect the requirement, dependency, or environment assumption instead of increasing attempts indefinitely.

## Download important artifacts

Do not rely on artifact retention as your permanent source-control strategy. Store important outputs in the system your project normally uses.

## Escalate systemic problems

If many unrelated builds are stuck or failing at the same stage, do not keep retrying user jobs. Operators should inspect queue, worker, provider, storage, validation, and incident health.

<Tip>
  The highest-quality FORGE workflow is usually **plan → small coherent change → validate → review → continue**, not “generate everything and hope the ZIP works.”
</Tip>