ForgeFX Working Guide

Token-Efficient AI Workflows

Use the smartest model where judgment matters, and let capable lower-cost models handle bounded execution. Optimize for completed work—not the cheapest first attempt.

Route by difficulty

Use medium or fast models for extraction, formatting, tests, mechanical edits, and routine implementation. Promote to frontier models for ambiguity, architecture, hard debugging, risk, and final review.

Plan, build, review

For substantial work: have a strong model define the approach, a cheaper model execute a bounded slice, then use the strong model for consequential review.

Keep context lean

Pass paths, constraints, acceptance checks, and focused excerpts. Avoid whole-repository dumps, giant logs, repeated file reads, and stale multi-day threads.

Measure finished work

A cheap half-job that needs three correction passes may cost more than one good pass. Compare matched tasks by total usage, retries, tool output, and verified completion.

The Three-Pass Pattern

Planner · strong model

Inspect the relevant code. Identify the approach, files, risks, and tests. Avoid editing until the plan is clear.

Builder · medium model

Implement one bounded unit. Preserve existing patterns, make the smallest complete change, and run focused checks.

Reviewer · strong model

Inspect the diff and test results. Look for missing work, regressions, and design risks. Apply only actionable fixes.

Cursor Setup

Use a fast model for autocomplete, a medium model for Ask/reconnaissance and routine Agent work, and a frontier model for architecture, hard debugging, and final review.

# .cursor/rules/project-workflow.mdc
- Inspect relevant files before editing; avoid broad scans.
- Make the smallest complete change and preserve existing patterns.
- Work only in the requested area unless a dependency requires otherwise.
- Run focused tests or checks after editing.
- Do not stop at a plan when implementation was requested.
- Summarize failures; do not dump full logs.
- Report changed files, checks run, and unresolved risks.

Context control: prefer explicit @Files over @Codebase, keep one objective per chat, and use .cursorignore for generated or bulky directories such as node_modules, dist, build, coverage, .next, maps, and generated assets.

Codex Setup

Keep AGENTS.md short and operational. Define the scope, relevant paths, acceptance criteria, focused tests, and the requirement to implement and verify—not merely propose.

Implement the approved plan in src/dashboard/.
Preserve the existing API and visual style.
Make the smallest complete change.
Run the focused dashboard tests.
Do not modify unrelated files or stop after scaffolding.
Report changed files, checks run, and remaining risks.

Default rule: medium model for routine implementation; frontier model for ambiguity, difficult reasoning, and final review. The goal is the lowest cost for verified completed work—not the lowest cost per individual call.