Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/prosusai/prism/claude-mdgit clone --depth 1 https://github.com/ProsusAI/prismWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/instructions/prosusai/prism/claude-md)<a href="https://agentmods.dev/instructions/prosusai/prism/claude-md"><img src="https://agentmods.dev/badge/instructions/prosusai/prism/claude-md.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.02987 | $0.02987 |
| Opus 5 | $0.01494 | $0.01494 |
| Sonnet 5 | $0.00597 | $0.00597 |
| Haiku 4.5 | $0.00299 | $0.00299 |
Grade A, and why
prism CLAUDE.md scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Runs shell commandslowCapability
Expected in a hook, worth knowing in a rule or an instructions file.
- **`subprocess.run()` not `os.system()`** — always use `capture_output=True, text=True, timeout=N`. How it starts
The opening of the file, as written. The whole thing — 168 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Project
Prism — knowledge layer for Claude Code and Cursor. Two things:
- Personal learning: hooks observe tool usage, a two-phase extraction pipeline (fast model proposes, strong model validates — via
claudeor Cursoragentper backend) converts patterns into engrams (living, decaying knowledge). Engrams flow back into the IDE via context files (.claude/prism.md,.cursor/rules/prism.mdc) and MCP tools. - Team skills: high-confidence engrams promote to skills published to a Cloudflare Worker-backed registry that teams query.
Hard constraints
- Hooks never block the IDE —
capture.sh(Claude Code) andcapture_cursor.sh(Cursor) must always exit 0; background spawns only. - Storage split — observations + sessions live in SQLite (
~/.prism/prism.db) via stdlibsqlite3. Engrams stay flat Markdown + YAML frontmatter; the engram index staysindex.json. No external DB, no ORM. - AI calls via IDE CLIs only —
claude --print(Claude Code) oragent -p(Cursor). Never import the Anthropic SDK or cursor-sdk. No API keys for extraction. Routed throughlib/agent_runner.py(resolve_backend+run_agent). - One extraction CLI per user — Claude-only needs
claude login; Cursor-only needsagent login. Not both unless you use both IDEs. - Custom YAML frontmatter parser — never import PyYAML. Split on
---, parsekey: valuelines. subprocess.run()notos.system()— always usecapture_output=True, text=True, timeout=N.- MCP stdout is protocol-only — any stray
print()in lib code corrupts the JSON-RPC stream. All logging to stderr. - Never read
.envfiles — config comes fromos.environonly. No dotenv parsing, no opening.envfiles.
Tech Stack
| Layer | Technology | Notes |
|---|---|---|
| Library / CLI | Python 3.12+ (stdlib only) | argparse, json, pathlib, subprocess, hashlib, fcntl |
| Hooks / installer | Bash (POSIX-compatible) | capture.sh / capture_cursor.sh → capture.py. Avoid Bash 4+ features (macOS ships 3.2) |
| AI calls | claude CLI or Cursor agent CLI |
Claude: haiku / sonnet. Cursor: cursor_models.fast / cursor_models.strong (defaults: composer-2.5[fast=false], claude-4.6-sonnet-medium). All via lib/agent_runner.py. |
| MCP server | Python stdio, JSON-RPC 2.0 | Protocol version 2025-03-26. Tools only, no resources/prompts |
| Storage | SQLite (stdlib sqlite3) + flat files |
prism.db = observations + sessions + observations_fts (FTS5); index.json = engram index; Markdown engrams |
| Registry API | Cloudflare Worker (TypeScript) | Wrangler 4.x, Node 22 LTS — for registry maintainers only, not end users |
| Registry backend | GitHub repo | Versioning, PRs, CI, and hosting for free. No database needed |
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 4d ago First seen · 168 lines · 2,987 tokens per session scan A 5ed6d3e143f5
prism CLAUDE.md is an instructions file published in the GitHub repository ProsusAI/prism (20 stars, last pushed 2mo ago), licensed Apache-2.0. It adds 2,987 tokens to every session, about $0.0149 per session on Opus 5. A static security scan graded it A with 1 finding (runs shell commands). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other instructions, from other repositories
vespertide AGENTS.md
AGENTS.md instructions for dev-five-git/vespertide, covering vespertide knowledge base, structure, where to look, data flow and conventions.
aeon CLAUDE.md
Instructions for aeonfun/aeon, covering aeon, how aeon works, strategy, voice and soul file hierarchy (read in this order).
mcp-structured-memory CLAUDE.md
Claude Code instructions for nmeierpolys/mcp-structured-memory, a project described as: Structured Memory MCP Server.
inkwell-memory CLAUDE.md
Instructions for veronchenko/inkwell-memory, covering claude.md — inkwellmemory, layout, multi-tenant mode (inkwellmultitenant=1), conventions and testing.
RNR-Enhanced-Cognee AGENTS.md
AGENTS.md instructions for vincentspereira/RNR-Enhanced-Cognee, covering rnr enhanced cognee implementation for codex, critical requirements, 1. ascii-only output (no unicode encoding), 2. dynamic categories (no hardcoded categories) and 3. standard memory mcp interface.
memory-mcp-1file AGENTS.md
Instructions for pomazanbohdan/memory-mcp-1file, covering ⛔ l0 invariants (never violate under any circumstance), detection heuristic, 🔀 phase transition routing (⛔ blocking), gate-0: phase identification & loading and 🧾 proof-of-load requirement (critical).