Progressive Clarity

Response design protocol

One answer. Three useful views.

Progressive Clarity is a response protocol for AI assistants. Every applicable answer renders At a glance, In context, and At depth together: a direct stopping point, new context for action, then purposeful specialist detail.

Choose an install See the example
01

At a glance

≤ 40 counted English prose words
02

In context

≤ 200 cumulative shallow words
03

At depth

Purposeful detail

All three views appear together in every applicable response. Each fact belongs in its earliest necessary view.

Install

Choose the package for your host.

ChatGPT, Claude, and Cursor packages are Advisory prompt-only guidance. They need no backend, but host selection and instruction following remain probabilistic. Optional local pc-core adds a mechanical gate for buffered Cursor or Claude Code output.

Private upload · v0.2.2

Claude Free Skill

Enable code execution and file creation, then open Customize → Skills → + → Create skill → Upload a skill. Select the Claude.ai archive and enable it.

The upload is private to the account. Automatic selection remains probabilistic; no live Claude conformance is claimed.

Download Free Skill Upload guide
Community candidate · v0.2.1

Claude plugin

Use the plugin ZIP locally with --plugin-dir or submit the public repository to the reviewed claude-community directory. It is not currently advertised as approved or catalog-installable.

After community publication only

/plugin marketplace add anthropics/claude-plugins-community
/plugin marketplace update claude-community
/plugin install progressive-clarity@claude-community
Download Claude plugin Community submission
Manual Skill · v0.2.1

Cursor

Extract the complete progressive-clarity folder into one project or user Skill root. Choose one root for the intended scope. Installation is manual; no catalog distribution is claimed.

.agents/skills/progressive-clarity/ or .cursor/skills/progressive-clarity/
Download manual Skill Cursor Skills guide

Two conformance surfaces

Prompt guidance or a local mechanical gate.

Neither surface needs a hosted backend. Advisory packages are instruction files. pc-core is a separate local Python 3.11+ process that invokes the existing Cursor or Claude Code CLI.

Advisory prompt-only

Skills guide the model.

The ChatGPT, Claude, and Cursor packages contain instructions, a license, and package metadata or static assets where needed. They install no hook, runtime, MCP server, service, analytics, or network dependency.

Discovery and automatic activation remain host-owned and probabilistic. A successful upload or install is not a behavioral guarantee.

Optional local enforcement

pc-core gates buffered output.

From a reviewed repository checkout, run the non-streaming wrapper with a trusted request and local state file. Replace cursor with claude-code for Claude. 

python3.11 -m pc_core wrap \
  --host cursor \
  --request request.json \
  --state conversation-state.json \
  --cwd /path/to/project
Read the local wrapper guide

Mechanically certified

What a wrapper pass guarantees

  • Trusted request, state, schema, and protocol versions, plus topic, branch, turn, and fact-count arithmetic.
  • Three non-empty views in order, canonical headings, and the deterministic 40/200 English prose budgets.
  • Declared correction, warning, quotation, fact-ID, reuse, catalog-coverage, and exact lexical-duplicate checks.
  • At most two candidates, invalid-output withholding, canonical rendering, and atomic local state replacement after a pass.

Still semantic and unverified

What a wrapper pass cannot guarantee

  • Factual accuracy, completeness, source quality, or correctness and completeness of a caller-supplied fact catalog.
  • Best fact allocation, paraphrased repetition, purposeful depth, or absence of a hidden semantic reversal.
  • Human safe stopping, warning necessity or sufficiency, legal effect, or any safety outcome.
  • Host activation, host-wide compatibility, concurrent state writes, or rollback of downstream host side effects.

Mechanical certification applies only to the complete candidate buffered, validated, rendered, and released by that wrapper. It does not apply to ordinary streamed host output or ChatGPT.

Package files are synchronized from the canonical source and checked byte-for-byte before this page serves them. Platform marks belong to their respective owners.

Scenario lab

The same facts, shaped for the reader.

Choose a domain and a reader. The facts stay fixed while the answer changes what it leads with. Every example renders all three views together and assigns each fact to one view.

Domain

Technology domain, CEO reader.

The scenarios, organizations, and figures are fictional.

Reader

CEO

Facts

An AI support assistant is due Friday. In evaluation, 18% of answers lack required citations; the launch gate is 5%. The fix is merged, and regression testing ends Thursday. Three pilot customers were promised access this week.

Question Should the company launch on Friday?

01

At a glance

Decision

Do not commit to Friday’s broad launch. In evaluation, 18% of answers lack required citations against a 5% launch gate, so release risk is known and material.

02

In context

Business context

Readiness: The fix is merged, but regression testing ends Thursday. The release decision should wait for that evidence.

Commitment: Three pilot customers were promised access this week. Customer success should explain that access remains conditional on the go/no-go result.

03

At depth

Decision gate

Go criteria: Approve wider release only after the frozen evaluation set passes and resolution quality remains stable. Block or label any answer that still lacks required evidence.

Rollout: If every gate passes, start with one pilot, review logs daily, limit traffic, and keep rollback ready. Expand after two clean evaluation runs and pilot feedback.

One response contract

All three views in every ordinary in-scope response.

There is no depth progression or alternate presentation state. Requests for more, less, or one named view still receive the same three sections, focused on the current request.

01 Required

Ordinary responses

Render At a glance, In context, and At depth exactly once and in that order. Each deeper section adds information rather than replaying earlier facts.

A brevity request shortens all three sections without removing them.

02 Focused

Follow-ups and exceptions

A substantive follow-up receives all three views focused on the selected branch. Exact outputs, transformations, complete procedures, controlling text, and narratives keep the structure their purpose requires.

When missing information prevents a complete or safe response, ask one focused clarification without view headings.

Correction continuity: A correction retracts and replaces the error and states the changed consequence or action first under At a glance, then continues with In context and At depth.

Response structure

One contract, three additive views.

Every supplied fact is allocated to its earliest necessary view. Deeper sections stay focused on information not already given.

  1. 01

    At a glance

    Gives the direct answer, decision-relevant consequence, material scope, and every indispensable caveat.

    ≤ 40 counted English prose words*
  2. 02

    In context

    Adds new rationale, scope, constraints, ownership, timing, controls, or next action.

    ≤ 200 cumulative shallow words*
  3. 03

    At depth

    Adds purposeful evidence, assumptions, alternatives, exceptions, procedures, and implementation guidance.

    No hard cap

* At a glance may use at most 40 counted non-warning English prose words. The 200-word budget is cumulative: it combines non-warning prose from At a glance and In context in this response, not across the conversation. Indispensable warnings and necessary correction repair text are counted separately. Facts stay in their earliest necessary view; a deeper view must add information, not repeat a fact merely to stand alone.

Protocol invariants

Four requirements define each stopping point.

  1. 01

    Complete

    Answer the immediate request at At a glance and every cumulative stopping point. Do not defer an essential fact.

  2. 02

    Accurate

    Later detail may qualify the answer, but it cannot silently make an earlier statement false.

  3. 03

    Additive

    Each deeper view contributes information not already supplied. An atomic fact belongs in one view.

  4. 04

    Safe to stop

    A reader can leave at any view without forming a materially wrong belief or taking a materially wrong action.

Inside the project

One contract, two delivery surfaces.

The repository keeps Advisory prompt-only packages separate from the optional local wrapper that can certify mechanically decidable properties before releasing a buffered answer.

  • Protocol

    SPEC.md defines the single response contract, budgets, fact allocation, corrections, safety precedence, and non-fit structures.

  • Agent Skill

    A portable prompt-only package contains the canonical instructions and license. It requires no backend and makes no deterministic activation or semantic-conformance claim.

  • Local mechanical enforcement

    pc-core validates versioned envelopes, heading order, budgets, declared fact mechanics, corrections, quotations, and explicit state transitions. Its certification ends at wrapper-buffered output.

  • Examples and evaluation

    Eight Advisory host cases cover ordinary responses, targeted follow-ups, correction, safety, clarification, and non-fit structures. Unit and property tests cover the local mechanics, packaging, hooks, and repository tooling.

progressive-clarity / project files

◆ SPEC.md normative protocol

└─ skills/ SKILL.md + LICENSE

└─ plugin manifests ChatGPT + Claude packages

└─ pc_core/ validator + renderer + wrapper

└─ adapters/ project hook templates

└─ tests/ unit + property tests

└─ evals/ E01 to E08 host cases

└─ examples/ worked conversations

└─ tools/ validation + packaging

Explore the source

Explore the project

Read the contract or open the live package.

The repository contains the full specification, skill files, worked examples, templates, acceptance cases, and tooling.

Read the specification Open ChatGPT plugin View repository