AI coding agent context engineering

ContextForge

Turn any repository into an AI-ready context pack — AGENTS.md, .cursorrules, CLAUDE.md, copilot-instructions.md and a spec, generated from your actual tree.

Context pack generator

Paste a file tree. Deterministic extraction runs first and always succeeds; the model-assisted pass is optional. Nothing you paste is stored.

Target agents
Fair use: 5 generations / 24h.

What is ContextForge?

ContextForge converts a repository snapshot into the instruction files that AI coding agents read before they touch your code. You paste a file tree; it returns AGENTS.md, .cursorrules, CLAUDE.md, .github/copilot-instructions.md, GEMINI.md and a spec document.

The pipeline has two stages. Stage one is deterministic: a parser reads the tree and any key files you provide and infers package manager, language, framework, test runner, CI, entry points and directory conventions. Stage two is optional and model-assisted: an LLM turns those facts into readable prose with constraints and examples.

Because stage one never depends on the model, the product degrades honestly. No API key means a rule-based draft, labelled as such — never a fake AI response.

How to use ContextForge

  1. Run `git ls-files` (or `find . -type f -not -path "*/node_modules/*"`) in your repo and paste the output.
  2. Optionally paste 1-3 key files: package.json, your entry point, or CI config. This upgrades inference to detection.
  3. Select the agents your team uses. Each target gets a file written in that tool's expected format.
  4. Generate. Review the Context Score and the list of items it could not determine.
  5. Copy the files into your repo, correct anything marked inferred, and commit.

Common mistakes when writing agent context files

Copying a template repo verbatimIt describes someone else's stack. The agent follows it confidently and wrongly.
Stating conventions the repo does not followWorse than an empty file — the agent trusts the stated rule over the code.
Omitting build/test/lint commandsThe agent guesses commands, fails, then improvises. Always state exact commands.
No forbidden-paths sectionNothing stops the agent editing migrations, generated code, or lockfiles.
Writing it once and never updatingContext files rot. Re-run the generator when the stack changes.
Assuming every agent reads every fileCursor reads .cursorrules; Claude Code reads CLAUDE.md; Copilot reads .github/copilot-instructions.md. Ship the right file per tool.

ContextForge vs writing it by hand

Starting pointExtracted from your real tree and config filesBlank file, copied from a template repo
Stack detectionPackage manager, framework, test runner, CI inferred from filesYou remember to write it
CoverageChecks for build/test/lint commands, naming, forbidden pathsWhatever you remembered that day
ConsistencySame checklist every run, scored 0-100Varies by author
Time to first draftSeconds30-90 minutes per repo

Do not commit the generated files unread. A context file that states conventions your repo does not follow is worse than no context file. ContextForge labels every inferred fact as inferred so you can correct it.

Pricing

Starter

$19/mo
Billed monthly
  • All 5 agent file formats
  • Context Score + open-questions list
  • 5 generations / 24h fair use
Choose Starter

Team

$49/mo
Billed monthly
  • Everything in Starter
  • 60 generations / 24h per workspace key
  • Shared conventions section
Choose Team

Enterprise Custom

Custom
Private deployment and bespoke rule sets
  • Self-hosted option
  • Invoice billing
  • Onboarding for your conventions
Contact us

Prices in USD. Subscriptions are billed by our merchant of record, Waffo. Cancel any time from your Waffo receipt link.

Frequently asked questions

What is ContextForge?

ContextForge is a generator that turns a snapshot of your repository into the instruction files coding agents read — AGENTS.md, .cursorrules, CLAUDE.md, .github/copilot-instructions.md — plus a machine-readable spec of the detected stack.

How does it work?

You paste a file tree and optionally a few key files. A deterministic extractor infers package manager, framework, language, test runner, CI system, entry points and directory conventions. An optional model-assisted pass turns those facts into prose. The extractor never depends on the model, so a missing API key degrades to a rule-based draft instead of failing silently.

Do you store my code?

No. The request is processed to produce a response and is not persisted. ContextForge has no database and no repository connector — it only reads what you paste into the form.

What happens if the AI is unavailable?

The endpoint fails closed. With no API key you get HTTP 503 with an explicit error code; if the upstream model call fails you get 502; if the fair-use quota is exhausted you get 429. You never receive a fabricated "AI-generated" result — a rule-based draft is always labelled source: rule-based.

Which agents are supported?

Cursor (.cursorrules), Claude Code (CLAUDE.md), OpenAI Codex (AGENTS.md), GitHub Copilot (.github/copilot-instructions.md), Gemini (GEMINI.md), and a generic AGENTS.md fallback that most tools read.

How much does ContextForge cost?

$19 per month for the Starter plan, $190 per year (equivalent to $15.83 per month, saving $38 — about two months free). Team is $49 per month or $490 per year. Enterprise is custom-priced via email.

Is the generated context file accurate?

It is an accurate restatement of what you pasted, and clearly-marked inference where it had to guess. Facts inferred from filenames are labelled inferred; facts read from package.json or CI config are labelled detected. Treat the output as a first draft to review.