Skip to content
/
Phase D Batch 8

Batch 08 (Phase D) — Context/PRP/Rule Templates & Linters

Batch 08 (Phase D) — Context/PRP/Rule Templates & Linters

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
ozzeron-prompt-pack-promptkit 6 skill-pack no no sequential no A
ai-context-templates 1 standalone-repo no no none no A
ai-context-linter 1 github-action no no none no A
sisyphusse-pro-pack 26 standalone-repo no no none no A
claude-codepro 1,728 npm-package yes (pilot) yes (Console, :4200) sequential yes (Claude + Codex) A
caliber 1,087 npm-package yes (@rely-ai/caliber) no sequential yes (fast + default) A
faf-cli 29 npm-package yes (faf) yes (faf show HTML) none no A
contextkit 169 standalone-repo no no sequential no A
agents-md 21,710 standalone-repo no no (Next.js website only) none no A
agent-md 80 standalone-repo no no none no A

Intra-batch patterns

This batch is unified by a single theme: how do you encode project-level AI governance in files? The 10 frameworks span the full spectrum from "one file named AGENTS.md" to "a rated-and-versioned YAML context format with WASM scoring kernel and IANA registration." Every framework except claude-codepro produces or standardizes at least one of: CLAUDE.md, AGENTS.md, AGENT.md, .cursorrules, or .faf files.

Three distinct implementation philosophies emerge:

  1. Format standards (AGENTS.md, AGENT.md): Define only the filename and location; all content is the developer's choice. Zero enforcement, maximum portability. AGENTS.md won this contest with 270x the star count of AGENT.md.

  2. Template libraries (ai-context-templates, sisyphusse-pro-pack, ozzeron-prompt-pack-promptkit): Opinionated starting-point files that developers copy into their projects. No installation, no commands — just content to plagiarize. Differ in scope: ai-context-templates covers CLAUDE.md + Cursor + PRP; sisyphusse-pro-pack covers CLAUDE.md + AGENTS.md only; ozzeron ships 23 skill-md files with inheritance chains.

  3. Active enforcement tools (ai-context-linter, caliber, faf-cli, claude-codepro): Validate or score context files, go beyond passive templates. ai-context-linter uses static CI rules; caliber uses a 100-point deterministic rubric; faf-cli uses a WASM scoring kernel with an IANA-registered format; claude-codepro runs hook-based validators on every file write.

The contextkit framework is an outlier: it is iOS/macOS Swift-domain-specific and is the only spec-lifecycle framework in the batch (4 phases: Spec.md → Tech.md → Steps.md → implementation). It has been archived since March 2026, superseded by PlanKit.

Most interesting finds

  1. caliber (rely-ai-org): The 100-point deterministic scoring rubric (6 categories, no LLM required) combined with the full hook automation loop (SessionStart → observe → UserPromptSubmit → observe → PostToolUse → check-sync → SessionEnd → finalize + learn) creates the most complete "context hygiene" lifecycle in the Phase D batch. The caliber generate command creates context files from codebase analysis; caliber score evaluates them; caliber learn extracts patterns from agent interactions. The three-command closed loop (generate → score → learn) is architecturally distinctive. At 17 contributors and 1,087 stars it is also the most community-supported native context-management framework.

  2. faf-cli (Wolfe-Jam): Two Zenodo-published research papers with DOIs back this project (DOI: 10.5281/zenodo.18251362 and 10.5281/zenodo.20348942), making it the only framework in the entire Phase D corpus that claims standards-track standing with academic citations. The IANA-registered MIME types (application/vnd.faf+yaml and application/vnd.fafm+yaml), .faf-dna birth certificate concept, and bi-directional sync between project.faf and CLAUDE.md via mtime-direction detection are engineering choices more typical of internet standards work than developer tooling. The faf show command (new in v6.7) renders project.faf as a self-contained HTML document — the closest thing to a human-readable spec export in the batch.

  3. AGENTS.md vs AGENT.md schism: The naming war between AGENTS.md (agentsmd/agents.md, 21,710 stars, 20 contributors, 23+ compatible tools) and AGENT.md (agentmd/agent.md, 80 stars, 1 contributor, 2 tools) is the defining competitive dynamic of this batch. AGENTS.md won by ecosystem-first adoption through Codex CLI; AGENT.md attempted a standards-body approach with RFC-style language and Sourcegraph domain ownership. Both specifications include migration commands that point at each other (AGENTS.md FAQ: mv AGENT.md AGENTS.md && ln -s AGENTS.md AGENT.md; AGENT.md spec: ln -s AGENTS.md AGENT.md for Gemini/Codex/OpenCode). As of analysis date AGENT.md is dormant (10 months no commits). This is a cleanly observable case study in format-war outcomes.

Items written as Tier C

None. All 10 frameworks had sufficient public material for full 11-file reports. ai-context-templates (1 star) and agent-md (80 stars, single-file repo) were candidates for Tier C but both had sufficient specification content for full treatment.

Cross-references discovered

  • ai-context-templates → ai-context-linter: Both are by MrDwarf7. The linter is explicitly described as the validation companion to the templates. ai-context-linter was written to enforce the quality standards that ai-context-templates defines informally.

  • sisyphusse-pro-pack → Karpathy 4-rule CLAUDE.md: The pro-pack explicitly builds on Karpathy's 4-rule minimal CLAUDE.md template, adding 8 rules. Relationship: extension/derivative.

  • contextkit → PlanKit: contextkit is archived and its README links to PlanKit as successor. Same author (FlineDev), lighter design, plugin-based rather than bash-installer-based.

  • contextkit → kiro (seed): The Spec.md / Tech.md / Steps.md phase structure in contextkit directly borrows from kiro's spec lifecycle pattern, adapted to run on standard Claude Code slash-commands.

  • AGENTS.md FAQ → AGENT.md: AGENTS.md's FAQ migration command explicitly converts AGENT.md files. AGENT.md specification's migration table acknowledges AGENTS.md dominance by listing a reverse-symlink for Gemini/Codex/OpenCode.

  • ai-context-linter → caliber (functional overlap): Both validate AI context files. ai-context-linter is CI/static; caliber is session-integrated/dynamic. They are complementary (CI vs runtime) rather than direct substitutes.

  • faf-cli → AGENTS.md (export): faf-cli's faf export command can generate AGENTS.md from project.faf. faf treats AGENTS.md as a downstream export format of the canonical .faf source.

  • claude-codepro → kiro (seed): The spec-plan skill in claude-codepro creates a spec-plan.md document with a phased spec lifecycle similar to kiro's requirements.md/design.md/tasks.md pattern.

Primitive count comparison

Framework Commands Skills Subagents Hooks MCP servers
claude-codepro 0 16 5 12 events 7
caliber 0 8 0 6 events 0
contextkit 20 0 9 0 0
ozzeron-prompt-pack 0 23 0 0 0
faf-cli 26 CLI cmds 0 0 0 0
ai-context-linter 0 0 0 1 (CI action) 0
sisyphusse-pro-pack 0 0 0 0 0
ai-context-templates 0 0 0 0 0
agents-md 0 0 0 0 0
agent-md 0 0 0 0 0

Format naming convention observations

  • AGENTS.md (AGENTS.md spec): plural, no version
  • AGENT.md (AGENT.md spec): singular, RFC framing
  • CLAUDE.md (sisyphusse, caliber): tool-specific, claude-code-locked
  • .cursorrules (ai-context-templates): tool-specific, cursor-locked
  • project.faf (faf-cli): YAML, tool-agnostic, IANA-registered
  • Context.md (contextkit): project-scoped, part of directory hierarchy
  • CALIBER_CONTEXT.md (caliber): generated artifact, machine-produced
← Phase D Batch 7
Phase D Batch 9 →