ultrainit.sh / prompts
CLAUDE.md8 starsChanged 7 months agoApache-2.0View on GitHub
14 markdown files, each the system prompt for one claude -p agent. Files are paired 1:1 with schemas in schemas/. The prompt defines task intent; the schema enforces output types. Every gather agent prompt requires file citations for every claim. No assertion without a real file path. This is enforced by explicit instructions: patterns.md includes "Do NOT invent patterns. Only report what you can prove exists." Any new gather prompt must include a similar evidence requirement. Both synthesizer.md and synthesizer-tooling.md…
In this repository22 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- .github/workflows/CLAUDE.mdCLAUDE.md
- lib/CLAUDE.mdCLAUDE.md
- schemas/CLAUDE.mdCLAUDE.md
- scripts/CLAUDE.mdCLAUDE.md
- tests/CLAUDE.mdCLAUDE.md
- add-bats-testSkill
- add-edge-testSkill
- add-gather-agentSkill
- add-integration-testSkill
- add-synthesis-fixtureSkill
- audit-prompt-schema-syncSkill
- bundle-and-verifySkill
- check-ci-locallySkill
- debug-agent-failureSkill
- fixture-syncSkill
- pre-pr-checklistSkill
- rebuild-test-imageSkill
- release-ultrainitSkill
- run-testsSkill
- tune-synthesis-contextSkill
- validate-generated-artifactSkill
The file
We split this CLAUDE.md into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
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
- Very specific: line 36
Every "Do NOT do X" must be paired with "Do Y instead." synthesizer.md explicitly forbids lone prohibitions.
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.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
prompts/ — Agent System Prompts
14 markdown files, each the system prompt for one claude -p agent. Files are paired 1:1 with schemas in schemas/. The prompt defines task intent; the schema enforces output types.
File Index
| File | Phase | Model | Purpose |
identity.md | 1 | haiku | Project name, languages, frameworks, monorepo detection |
commands.md | 1 | haiku | Build, test, lint, format commands from all config files |
git-forensics.md | 1 | sonnet | Hotspots, temporal coupling, bug density, ownership |
patterns.md | 1 | sonnet | Architectural patterns, error handling, naming conventions |
tooling.md | 1 | haiku | Linters, formatters, pre-commit hooks |
docs-scanner.md | 1 | haiku | Existing docs inventory and coverage gaps |
security-scan.md | 1 | haiku | Files needing protection, safe alternatives |
structure-scout.md | 1 | sonnet | Directory map with roles and priority tiers |
module-analyzer.md | 1 (Stage 2) | sonnet | Deep-dive per directory (spawned once per important dir) |
domain-researcher.md | 3 | sonnet | Business domain terminology and framework best practices |
mcp-discoverer.md | 3 | sonnet | MCP server recommendations from registry + GitHub |
synthesizer-docs.md | 4 Pass 1 | sonnet[1m] | Generate CLAUDE.md files from all findings |
synthesizer-tooling.md | 4 Pass 2 | sonnet[1m] | Generate skills, hooks, subagents, MCP config |
synthesizer.md | reference | — | Master/reference prompt covering both passes (not used directly in pipeline) |
Critical Rules
Evidence Mandatory
Every gather agent prompt requires file citations for every claim. No assertion without a real file path. This is enforced by explicit instructions: patterns.md includes "Do NOT invent patterns. Only report what you can prove exists." Any new gather prompt must include a similar evidence requirement.
Portability Test (Synthesizers)
Both synthesizer.md and synthesizer-tooling.md end with a checklist that includes the Portability Test: "Any skill that could be dropped unchanged into an unrelated project has failed." Every skill must contain ≥3 codebase-specific file references to pass. This is the most important quality gate in the entire system.
Prohibition + Alternative Pairing
Every "Do NOT do X" must be paired with "Do Y instead." synthesizer.md explicitly forbids lone prohibitions. This applies to CLAUDE.md content, skills, and subagent instructions. Do not add prohibitions to synthesizer outputs without alternatives.
Generic Phrase Ban
These phrases are explicitly banned in synthesizer.md and must never appear in generated CLAUDE.md or skill content: "best practice", "clean code", "SOLID principles", "maintainable", "readable", "scalable", "well-structured", "production-ready", "industry standard". Delete on sight.
YAML Angle-Bracket Prohibition
synthesizer-tooling.md explicitly forbids < and > in YAML frontmatter fields. This prohibition is absent from synthesizer.md, even though it also generates skill YAML. When editing synthesizer.md, add: "NEVER use angle brackets (< or >) in YAML frontmatter or description fields — they break YAML parsing."
Schema Coupling
Prompts describe output field names in prose; schemas enforce types. The two must stay in sync manually — there is no automated cross-validation. When modifying a prompt's output section, always update the paired schema in schemas/ in the same commit.
module-analyzer.md receives the directory path as context injected into the prompt string at call time in lib/gather.sh, not via a system prompt variable. If refactoring the invocation, the path must be interpolated before the claude -p call.
Complexity vs Phase
Gather agents: 13–55 lines (narrow, single output type, cost-sensitive — many run in parallel). Synthesizers: 101–295 lines (multi-artifact, quality checklist, embedded template structures). This ratio is intentional. Never add quality checklists to gather agents.
Known Issue: Three Overlapping Synthesizer Prompts
synthesizer.md is the comprehensive reference. The actual pipeline uses synthesizer-docs.md (Pass 1) then synthesizer-tooling.md (Pass 2). Using synthesizer.md as a drop-in replacement for both passes would skip the Pass 1→Pass 2 dependency (Pass 2 needs Pass 1's CLAUDE.md as source of truth). Do not conflate them.
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.

