Skip to content

AkashAi7/stenographer-mode

Repository files navigation

Steno Mode — GitHub Copilot Prompt Pack & Custom Agent

Stenographer Mode for GitHub Copilot, Claude Code, Kiro, Cursor & More. Structured shorthand that cuts ~40% of AI response tokens while keeping every code literal, command, and path exactly intact.

npm License: MIT Benchmark Platforms


Watch the Steno Mode demo

▶ Watch the demo · Open interactive benchmark · GitHub repo


What It Does

Steno Mode is a prompt-driven response style for AI coding assistants. It compresses answers through consistent shorthand rules, not random truncation.

Technique Example
Stable abbreviations cfg, auth, deps, req, resp, impl, ctx
Symbolic chains A -> B -> C instead of prose transitions
Filler removal Drops pleasantries and repeated framing
Literal protection Code, paths, commands, errors stay verbatim

Quick Start

If you want the shortest path:

  1. Install it in your repo.
  2. Use /steno brief in Copilot Chat, Agent mode, or any custom agent.
  3. Keep brief for normal use, switch to lite when you want more prose.

Best default: install project-wide so Steno applies across Ask, Edit, Agent, chat agent mode, and custom agents in that workspace.


Benchmark

Real numbers from 50+ samples across 7 categories (debugging, code-review, architecture, docs, onboarding, ambiguous, stakeholder):

Baseline Caveman Steno
Tokens (retry loop Q) 52 27 20
Avg reduction vs baseline ~48% ~40%
Readability win rate 86% over caveman
Literal accuracy 100% 100% 100%

Failure cases are included in the corpus. Steno is not universal — see Honest Scope.


Quick Comparison

Prompt: Why does this API retry loop never stop?

Baseline  The retry loop never stops because the retry counter is stored inside
          the request handler, so it resets to zero on every new attempt.
          Move the counter to state that survives across attempts.           [52 tok]

Caveman   Retry counter inside request handler. Resets each retry. Terminal
          condition never hit. Move counter outside.                         [27 tok]

Steno     Retry ctr lives inside req handler -> resets each attempt ->
          no terminal hit. Persist ctr across attempts.                      [20 tok]

Prompt: Review this caching change.

Baseline  This change improves cache hit rate, but it also introduces a stale
          data risk because invalidation only occurs on create and not on
          update or delete.

Steno     Hit rate up, but cache invalidation only covers create ->
          stale reads on update/delete paths.

Install

Recommended

Use this if you want Steno available in the current repo:

# Current repo only (.github/prompts command set + .github/agents + .github/copilot-instructions.md)
npx --yes github:AkashAi7/stenographer-mode install --scope project

Optional

Use this if you want the user profile setup instead:

# VS Code user profile (prompt command set + skill + agent)
npx --yes github:AkashAi7/stenographer-mode install --scope user

# Global CLI
npm install -g github:AkashAi7/stenographer-mode

If you already cloned the repo:

npm install
npm run install:user    # or install:project

install:project also writes .github/copilot-instructions.md, which is what makes Steno apply across Ask, Edit, Agent, chat agent mode, and custom agents in that workspace.

Remove it:

npm run uninstall:user

Usage

VS Code Copilot Chat

/steno Why does this test fail intermittently?

Common choices:

  • /steno brief -> default, easiest to read
  • /steno lite -> more prose
  • /steno court -> denser shorthand
  • /steno machine -> maximum compression

Switch compression levels inline:

Command Style
/steno lite Tight professional prose
/steno brief Default shorthand (recommended)
/steno court Dense expert shorthand
/steno machine Maximum compression

Persistent mode: say Steno Mode once, or install the project instructions file. The skill keeps it active across Ask, Edit, Agent, chat agent mode, and all custom agents until you say stop steno.

Compress a context file:

/steno-compress temporary   # returns shorthand in chat only
/steno-compress permanent   # rewrites the target file in-place

Bulk convert existing .github files:

/steno-convert-github preview
/steno-convert-github apply

Deterministic CLI backend used by this command:

node ./scripts/steno-mode.mjs convert-github --mode preview --project-dir .
node ./scripts/steno-mode.mjs convert-github --mode apply --project-dir .

NPM shortcuts:

npm run convert:github:preview
npm run convert:github:apply

Examples for .github files:

/steno-compress temporary .github/copilot-instructions.md
/steno-compress temporary .github/skills/steno/SKILL.md
/steno-compress permanent .github/agents/steno.agent.md

Use temporary to preview compression first, then run permanent on the same file when the output looks good.

VS Code Agent Mode

Pick the Steno agent from the agent picker, then ask normally:

Review this diff for regressions.
Use court for terse progress updates.

You can also use Steno as a mode on top of any other agent.

How to use it with any agent:

  • Install the project scope so .github/copilot-instructions.md is present.
  • Switch to any agent you want to use.
  • Say Steno Mode once, then continue with your normal task.
  • To stop, say stop steno or normal mode.

Example flow with another agent:

Switch to your preferred agent.
Steno Mode
Review this diff for regressions.
Keep updates brief.

What this does:

  • the selected agent keeps its own expertise and tools
  • Steno changes the response style to shorthand-first compression
  • this works across Ask, Edit, Agent mode, and custom agents in the workspace when project instructions are installed

Easy Mental Model

  • project install -> Steno rules are written into the repo and apply across every Copilot mode, agent mode, and custom agent in that workspace.
  • user install -> Steno is available from your VS Code profile.
  • Steno agent -> direct shortcut when you want the behavior on demand.
  • Steno Mode + any other agent -> keep that agent's specialization, but make its visible responses follow Steno style.

Platforms

Platform How to activate
VS Code Copilot npx install → /steno
Claude Code Paste packs/claude/system.txt into system prompt
Kiro Paste packs/cursor/rules.txt into agent instructions
ChatGPT Paste packs/chatgpt/custom-instructions.txt
Cursor Paste packs/cursor/rules.txt into rules
Windsurf Paste packs/cursor/rules.txt into rules
Aider --system-prompt flag or .aider.conf.yml
Sourcegraph Cody Paste into Cody custom instructions

Honest Scope

Steno works well in some contexts and poorly in others.

Works well:

  • Code review comments
  • Bug explanations and debugging Q&A
  • Architecture summaries
  • API and config documentation

Does not work well:

  • Onboarding and tutorials (beginners need prose)
  • Stakeholder communication (executives expect full sentences)
  • Empathetic responses
  • Teaching new concepts where analogies need space

Compression Levels

Level Description
lite Tight professional prose. Full sentences mostly intact.
brief Default. Shorthand + symbols + compact phrasing.
court Dense expert shorthand. Fragments allowed.
machine Maximum compression. Expert use only.

Benchmarking

npm install
npm run benchmark

Outputs benchmarks/latest.json and demo/benchmark-data.js using exact token counts via gpt-tokenizer.


Repo Structure

.github/
  agents/steno.agent.md          VS Code custom agent
  skills/steno/SKILL.md          Copilot persistent skill
  skills/caveman/SKILL.md        Caveman comparison skill
bundles/vscode/
  steno.prompt.md                VS Code prompt bundle
  steno-compress.prompt.md       Context compression command
  steno-convert-github.prompt.md Bulk .github conversion command
.github/copilot-instructions.md  Project-wide Steno instructions
packs/
  claude/system.txt              Claude / Claude Code pack
  chatgpt/custom-instructions.txt
  cursor/rules.txt               Cursor / Kiro / Windsurf pack
benchmarks/latest.json           Benchmark results
scripts/steno-mode.mjs           Cross-platform installer CLI

Repository

About

Shorthand-first token compression product with VS Code prompt bundles, exact token benchmarking, and cross-platform starter packs.

Resources

License

Contributing

Security policy

Stars

6 stars

Watchers

0 watching

Forks

Packages

 
 
 

Contributors