Agent task harness for coding agents

Preflight the task.
Loop until done.

Turn rough requests into harness-aware tasks with routing, skills, and judge commands baked in—then close the loop with real shell tests, not LLM self-report. One @im call → agent implements → run judges → fix & repeat until Loop complete. Works in VS Code, Cursor, Claude Code, Cowork, Codex, and terminal.

920 skills wired into @im - local-first, no API key required

agents + terminal Marketplace

curl -fsSL https://improv.sh/install.sh | bash

chrome extension

Add to Chrome

Or sideload:

curl -fsSL https://improv.sh/install-browser.sh | bash

dashboard — Cursor

loader.ts

DataTable.tsx

dashboard› src›

loader.ts

main Ln 12, Col 28 · Spaces: 2 · UTF-8 · TypeScript

// Extension

Built for developers.
Works on any platform.

IDE & terminal

Enhance prompts in seconds with @im or /im. Diff, apply, run judges, loop until exit code 0 — without leaving your editor.

VS Marketplace →

Chrome extension

Same local engine on web chat. Score, compile, and apply structured tasks before you send — no tab switching.

Add to Chrome →

// Compare

Stop paying agents to rediscover the task.

The value is not prettier wording. @im compiles repo context, constraints, deliverables, acceptance criteria, and shell validation into one spec — so the agent implements on turn one instead of interviewing you.

Raw prompt ~900 discovery tokens

Median 2 clarification turns before useful implementation.

With @im ~454 prep tokens

Median compiled spec from @im --json --provider local.

You save ~613 tokens median

Median savings (IQR ~444–866).

without @im

user: Fix login bug

agent turn 1:
- Which login flow is broken?
- What error do you see?
- Should I add tests?

user: OAuth callback. 500 after redirect. Yes.

agent turn 2:
- Which auth provider?
- Cookie or JWT session?
- What test command should I run?

~1.5k–6k discovery tokens before useful work.

with @im

user: @im Fix login bug

## Objective
Fix the login bug in the OAuth callback flow so sign-in completes without a 500.

## Context
- `src/auth/oauth.ts` handles callback + session mint
- Repo uses cookie sessions; `npm test -- auth` exists
- Original request: "Fix login bug"

## Constraints
- Preserve existing public auth API
- Scoped change — no drive-by refactors
- Reuse current test framework and fixtures

## Deliverables
- Root-cause fix in the callback path
- Regression test for redirect → session
- Summary of changed files + validation results

## Acceptance Criteria
- [ ] Login succeeds for the affected flow
- [ ] Session behavior does not regress
- [ ] Expired/invalid callback covered by a test

## Validation Steps
- Run `npm test -- auth/oauth`
- Manually verify login → redirect → dashboard
- Exit 0 on targeted auth suite before done

~200–700 prep tokens → implement on turn 1.

// Loop

Infrastructure for the full agent loop.

Most tools stop at rewriting the prompt. improv scopes, routes, and remembers the task—then closes the loop with real shell judges. Copy loop commands →

Loading library stats…

autoresearch loop

Self-improving library

Daily search → score → auto-adopt → corpus merge. Skills and harnesses powering @im grow every day.

repo-aware context

Your stack, preloaded

Detects package scripts, test commands, and agent instruction files from the workspace.

task memory

Agents forget. Tasks don't.

Lifecycle id, status tracking, and specs saved under .improv/tasks/.

external judges

Exit code 0, not vibes

Done means your repo's real commands pass—npm test, typecheck, build.

execution discipline

Ralph loops, slice by slice

Slice-by-slice checklists—inspect, implement, verify, report.

// Skills

An agent task harness, not a thesaurus.

Routing, skills, judge commands, and task memory in one @im call — an operating layer for agent work. Browse example prompts →

Surface awarenessRoutes for VS Code, Cursor, Claude Code, Cowork, Codex, and shell—with the right syntax for each.

scroll-worldScroll-scrubbed fly-through landing pages — Higgsfield CLI for production clips, frame-matched seams, portable scrub engine. Planning works without Higgsfield; clip generation needs install + auth.

codex-firstOn Claude Code: freeze the spec, delegate implementation to Codex CLI, verify the diff yourself—Claude thinks, Codex types.

Agent-aware rewritesSurface-specific discovery, patch discipline, validation commands, and summary expectations.

Customizable skillRewrite logic in canonical SKILL.md—read, edit, sync across harnesses.

Multi-harness syncInstall writes skills to .claude/skills, .cursor/skills, .agents/skills, .github/skills.

Clarify over assumeUp to 5 clarifying questions when missing context would change the implementation.

// How it works

3 steps. Under 10 seconds.

1
Write your prompt
Type as usual, then hit /im or @im.
2
Preflight compiles
Routing, skills, acceptance criteria, and judges — in one spec.
3
Loop until done
Agent implements. Run judges. Repeat until complete.

// Surfaces

Command routes for each environment

The same engine through the shortcuts developers already reach for.

VS Code

Sidebar panel with diff, Apply, Run judges, and @im in Chat.

@im Fix login bug

Cursor

Slash commands — 8 task template variants.

/im Fix login bug

Claude Code

Global /im synced at install. Routes codex-first for implementation work.

/im Refactor dashboard loader

Claude Cowork

Judge loop, tasks, templates via cowork-im.

/one-click-improve:im Fix login bug

Terminal

Pipe specs into Claude Code, Codex, or CI.

@im "Fix login bug"

Codex

Copy or pipe compiled specs.

@im "Add auth tests" | pbcopy

Web Chat

@im on ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek, Copilot, Poe.

Add to Chrome → #browser-install

Templates: /im-bugfix /im-refactor /im-a11y /im-safe /im-fast /im-deep /im-test-heavy

// Providers

Local for structure. Model-backed for ambiguity.

local is the default: fast, offline, deterministic. auto and model providers are for messy, domain-heavy asks. VS Code keys in SecretStorage; terminal keys in ~/.config/one-click-improve/env.

localSpeed, offline, deterministic task structure

autoFirst configured model key, otherwise local

openai / anthropicAmbiguous or product-specific rewrites

compatibleOpenRouter, Groq, Mistral, DeepSeek, Gemini, xAI

Pipe-friendly CLI

@im --json emits schema v1 with analysis, compiled spec, token savings, and memoryPath.

@im --json @im --tasks @im gates @im --template bugfix

// Install

Install once, improve prompts everywhere

Install from the VS Marketplace, one shell command for agents and terminal, or sideload the Chrome extension.

Agents & terminal

Detects VS Code and Cursor, installs the extension, creates @im, @harness, cowork-im shortcuts, syncs skill library. Or install from VS Marketplace.

zsh

curl -fsSL https://improv.sh/install.sh | bash

code --install-extension improv.one-click-improve

Optional: add .improv.md or .improv/config.json at your repo root for team stack, test commands, and security rules — @im loads it automatically.

chatgpt · claude · gemini · perplexity · grok · deepseek · copilot · poe

Chrome extension for web chat

Score, improve, and apply agent preflight tasks inside web chat — same local @im engine.