prism

Prism

Instruments over personas. A Claude Code plugin for building domain-expert analysis agents that actually reason like experts — not by pretending to be one, but by loading the structured analytical instruments experts actually use.

v0.5.5 ships a library of 742 instrument files across 5 classes and 62 domains, all indexed by a single catalog.yml triage file.

The problem with persona prompts

Telling an LLM "You are a senior security engineer" changes its tone but not its reasoning depth. It will write more assertive sentences and sprinkle in domain vocabulary, but the underlying analysis remains ad-hoc. Ask it to review an API and you'll get a confident-sounding list of suggestions that misses the same things an untrained reviewer would miss.

Real security engineers don't run on vibes. They walk checklists: STRIDE for threat modeling, OWASP Top 10 for API review, CVSS for severity. These checklists are the reason their analysis is deep — they force coverage, they surface blind spots, and they produce findings that compare across reviews.

The same is true in every domain. Curriculum reviewers use Bloom's Taxonomy and achievement-standard alignment. Strategists use Porter's Five Forces and pre-mortems. Decision scientists use Prospect Theory and Occam's Razor. Cultural critics read through Marxist and Foucauldian stances.

Prism treats these instruments as first-class, reusable assets. You build agents by composing instruments, not by writing adjective-heavy persona prompts.

The 5 classes

In v0.1, Prism called every instrument a "lens" — which was imprecise. Real instruments come in different shapes, and flattening them all into one template loses information. v0.2 introduced an explicit 5-class taxonomy; v0.3 scales it out to 657 files. See CLASSES.md for the full reference.

Class	What it is	Examples
Lens (렌즈)	Structured executable analytical procedure with input, steps, output format, and confidence signal.	STRIDE, OWASP API Top 10, CVSS, Bloom's Taxonomy, First Principles, Pre-mortem
Frame (분류 프레임)	Taxonomy or classification matrix — categories with discriminating criteria, no built-in procedure.	Cynefin, Kano Model, BCG Matrix, Webb's DOK, MoSCoW
Model (이론 모델)	Descriptive or predictive theoretical model — variables, relationships, predictions.	Porter's Five Forces, Prospect Theory, Nash Equilibrium, Lotka-Volterra
Stance (비판적 관점)	Interpretive position — commitments about what is worth looking for.	Marxist criticism, Foucauldian power/knowledge, Deconstruction, Bourdieu's Habitus
Heuristic (원리 / 격언)	Single rule of thumb — too small to be a lens, used as a check.	Occam's Razor, Hanlon's Razor, Chesterton's Fence, Hickam's Dictum

The strict definition of a lens is specific: it must have a defined input type, an executable procedure, a structured output format, AND a confidence evaluation. If any of the four is missing, the instrument belongs to one of the four sister classes.

Agent workflow

Agents run a 7-step pipeline. Simple agents use only steps 1, 3, and 7; richer agents enable the optional middle steps as needed.

1. Triage          — read catalog.yml; select items of any class
2. Classify        — apply FRAMES to sort the input               [optional]
3. Analyze         — apply LENSES in sequence                     [core]
4. Model-interpret — instantiate MODELS on input or findings      [optional]
5. Critical-read   — apply STANCES to surface hidden structures   [optional]
6. Sanity-gate     — run HEURISTIC bundles as final checks        [optional]
7. Synthesize      — cross-check across all items, prioritize

Each class maps to exactly one workflow step. The agent config declares which items to load and what usage tag each has (always, when-relevant, on-request).

Core ideas

• Instruments > personas. An agent's analytical depth lives in the instrument files. The persona is a one-line framing; the instruments do the work.
• Instruments are reusable. Once written, the same stride.md powers a security-reviewer agent, a DevSecOps pipeline agent, and a threat-modeling workshop facilitator.
• Catalog first, load on demand. Every item lives in catalog.yml at the repo root as a triage index. Agents read the catalog first, select what applies, and only then load the full files. This keeps context small.
• Agents stay thin. An agent config is ~40 lines: a one-sentence persona, per-class lists of item references with usage tags, and the workflow steps the agent actually uses.

Installation and usage