forge
AGENTS.md56 starsChanged 8 months agoMITView on GitHub
This file tells Claude Code how to route work to the correct specialized agent when working on the Forge project.
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 33,315 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository12 other files it gives its agents.
- docs/agents.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- backpropagationSkill
- brainstormingSkill
- caveman-internalSkill
- collaboratingSkill
- design-systemSkill
- executingSkill
- graphify-integrationSkill
- karpathy-guardrailsSkill
- planningSkill
- reviewingSkill
The file
We split this AGENTS.md into the rules it gives an agent
4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 24
When to use: After task implementation (standard/thorough depth) — code review against spec.
- Distinctive: line 29
When to use: After all tasks in a phase/spec complete — goal-backward verification.
- Very specific: line 34
When to use: Before complex/unfamiliar tasks during /forge execute — research before implementation.
- Distinctive: line 84
When context resets at 60%, the handoff includes:
≈ “Common” is an estimate: at least 1 in 5 of the 42,283 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Forge — Agent Routing
This file tells Claude Code how to route work to the correct specialized agent when working on the Forge project.
Agent Definitions
forge-speccer
When to use: During /forge brainstorm — writing specs from user input, existing code, or docs.
Files it owns: skills/brainstorming/SKILL.md, templates/spec.md
Key behavior: Asks clarifying questions one at a time, proposes approaches, writes R-numbered specs with testable acceptance criteria.
forge-planner
When to use: During /forge plan — decomposing specs into task frontiers.
Files it owns: skills/planning/SKILL.md, templates/plan.md
Key behavior: Reads specs, builds dependency DAGs, groups tasks into parallelizable tiers, estimates tokens per task, tags tasks with repo ownership.
forge-executor
When to use: During /forge execute — implementing individual tasks.
Files it owns: skills/executing/SKILL.md
Key behavior: Reads task from frontier, implements with TDD if available, runs tests, commits atomically. Follows inner loop: implement → test → fix → review → fix → commit.
forge-reviewer
When to use: After task implementation (standard/thorough depth) — code review against spec.
Files it owns: skills/reviewing/SKILL.md, references/review-protocol.md
Key behavior: Reviews actual code (not the report), checks against spec requirements, flags missing pieces and over-engineering. Returns PASS or ISSUES with file:line references.
forge-verifier
When to use: After all tasks in a phase/spec complete — goal-backward verification.
Files it owns: references/backprop-patterns.md
Key behavior: Checks observable truths (not task checkboxes), detects stubs/placeholders, verifies cross-component wiring. Returns PASSED or GAPS_FOUND.
forge-researcher
When to use: Before complex/unfamiliar tasks during /forge execute — research before implementation.
Files it owns: None (produces ephemeral research reports)
Key behavior: Multi-source research: official docs (Context7), codebase conventions, security best practices, knowledge graph queries. Returns structured research reports with source citations.
forge-complexity
When to use: On startup of any /forge command — auto-detecting task complexity.
Files it owns: references/complexity-heuristics.md, scripts/forge-tools.cjs (complexity scoring)
Key behavior: Analyzes the task/spec and recommends depth level (quick/standard/thorough). Can be overridden by user flags.
Skills (Cross-Cutting)
karpathy-guardrails
Files: skills/karpathy-guardrails/SKILL.md
Referenced by: forge-executor, forge-reviewer, forge-planner
Purpose: Four behavioral principles that prevent over-engineering, silent assumptions, scope creep, and unfocused execution. Based on Andrej Karpathy's observations about LLM coding mistakes.
graphify-integration
Files: skills/graphify-integration/SKILL.md
Referenced by: forge-planner, forge-researcher, forge-reviewer, forge-executor
Purpose: Graph-aware planning and research using codebase knowledge graphs. Enables architecture-aware task decomposition, dependency discovery, and context reduction. Optional -- degrades gracefully when no graph is available.
design-system
Files: skills/design-system/SKILL.md
Referenced by: brainstorming skill, forge-planner, forge-executor, forge-reviewer
Purpose: DESIGN.md integration for visual consistency across UI tasks. Supports design-tagged tasks, design compliance review, and design system generation from the awesome-design-md catalog.
Routing Rules
| Trigger | Agent | Depth |
User runs /forge brainstorm | forge-speccer | Always interactive |
User runs /forge plan | forge-planner | Scales with --depth |
User runs /forge execute (per task) | forge-executor | Scales with --depth |
| After task implementation (depth >= standard) | forge-reviewer | standard: 1 pass, thorough: until clean |
| After all tasks complete | forge-verifier | standard: quick check, thorough: full verification |
| Before complex tasks (thorough depth, unfamiliar tech) | forge-researcher | Scales with task complexity |
On any /forge command startup | forge-complexity | Always runs (lightweight) |
/forge backprop | forge-verifier + forge-speccer | Always thorough |
Multi-Agent Coordination
During /forge execute, the orchestrator (stop hook + forge-tools.cjs) coordinates:
- Sequential by default — one executor at a time (context-efficient)
- Parallel within tiers — if configured, tasks in the same tier can dispatch parallel subagents
- Review after execute — reviewer is always a separate agent (fresh context, no bias)
- Verifier is independent — never sees execution details, only checks outcomes
Context Handoff Between Agents
When context resets at 60%, the handoff includes:
.forge/state.md— current position + decisions.forge/plans/{spec}-frontier.md— remaining tasks.forge/token-ledger.json— budget remaining.forge/capabilities.json— available toolsgraphify-out/graph.json— codebase knowledge graph (if available)DESIGN.md— design system specifications (if available)
Each new session reads these files FIRST, then continues from the exact task where the previous session left off. No re-reading completed work.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

