one-click-improve

fela@agents:~$ @im "Fix login bug"

Prompt compiler for AI coding agents

One-Click Improve

Turn rough requests into harness-aware tasks before they reach Codex, Claude Code, Cursor, Cowork, or VS Code. It adds scope, acceptance criteria, and validation—then @im --verify when the agent says it's done.

@im terminal, Codex, VS Code /im Cursor and Claude Code @harness load surface rules
root@one-click-improve: ~/install
curl -fsSL https://improv.sh/install.sh | bash
No key required Best with your key Local + model-backed

Self-improving library · loading skills and harnesses…

same prompt, different outcome

Stop paying agents to rediscover the task.

The value is not prettier wording. It is whether the first model turn receives scope, harness context, acceptance criteria, and validation instead of a vague request.

raw prompt ~900 discovery tokens

Median 2 clarification turns before the agent has enough scope to start useful implementation.

with one-click improve ~454 prep tokens

Median compiled spec size from @im --json --provider local across 30 rough prompts.

estimated savings ~613 tokens

Median savings (IQR ~444–866). Regenerate with npm run benchmark:savings.

without one-click improve

user: Fix login bug

agent first response:
- Which login flow is broken?
- What error do you see?
- What files should I inspect?
- Should I add tests?

cost:
The session starts with discovery. The agent has no target flow, no done condition, no validation plan, and no surface-specific instructions.

token burn estimate:
~1.5k-6k tokens before useful work.

with one-click improve

user: @im Fix login bug

one-click output:
## Objective
Fix the login bug in the affected auth flow.

## Harness
Use the active route for this surface: VS Code @im, Cursor /im, Claude Code /im, Cowork /one-click-improve:im, terminal @im, or Codex-ready copy.

## Context
Use the selected workspace, current editor/chat surface, and bundled improve-prompt skill. Confirm the failing login path before editing.

## Constraints
- Preserve auth API and session behavior unless the bug requires a focused change.
- Avoid unrelated account or UI refactors.

## Deliverables
- Code fix.
- Tests for the failing login path.
- Validation notes for valid and invalid credentials.

## Acceptance Criteria
- Login succeeds for the affected flow.
- Existing logout/session behavior does not regress.
- The edge case is covered by a test.

## Validation
Run targeted auth tests and record the manual login check.

token burn estimate:
~200-700 prep tokens before useful work.

preflight → execute → verify

Infrastructure for the full agent loop.

Most tools stop at rewriting the prompt. One-Click Improve scopes the task before the agent starts, remembers it when context runs out, and checks the final report against criteria you already agreed on.

autoresearch loop Self-improving library

Daily search → score → auto-adopt → corpus merge. The live counter shows total skills and harnesses powering @im, what was added today, and the latest absorbed libraries—growing every day without manual curation.

repo-aware context Your stack, preloaded

Detects package scripts, test commands, and agent instruction files from the workspace so the first model turn already knows how to run and validate work.

task memory Agents forget. Tasks don't.

Every compiled task gets a lifecycle id, tracks status (generated → applied → sent → verified), and is saved under .improv/tasks/ with acceptance criteria and progress.

post-agent verify Know when you're done

Paste the agent's final report into the panel or run @im --verify to check evidence against the preflight task—not vibes.

weakness analysis 0–100 readiness score

Scores every prompt on 8 criteria: clarity, specificity, missing context, constraints, success criteria, testing, ambiguity, and architecture—then optionally asks up to 5 clarifying questions instead of inventing context.

panel review Diff, edit, Apply, Copy

Review original vs improved side-by-side in VS Code, edit inline, then Apply to the editor or Copy into chat—after optional clarifying questions when speed matters less than scope.

execution discipline Ralph loops, slice by slice

Compiled specs include slice-by-slice execution checklists—inspect, implement, verify, report—so agents finish one scoped change before starting the next.

agent-prep loop

What happens before your agent sees the prompt

One-Click Improve compiles intent into a task envelope: surface route, repo context, agent-aware discovery rules, patch discipline, deliverables, acceptance criteria, validation commands, and Ralph-style slice-by-slice execution.

01

$ invoke

Panel, selection, VS Code @im, Cursor /im, Claude Code, Cowork, or terminal @im.

02

$ inspect

Score readiness, load repo context, and detect missing scope, files, constraints, tests, and done conditions.

03

$ compile

Run the bundled skill and rewrite the request into an executable task with validation steps.

04

$ persist

Save to .improv/tasks/ with a lifecycle id so agents can resume without losing scope.

05

$ route

Send or copy the result in the command style that fits Cursor, VS Code, Cowork, Codex, or shell.

06

$ verify

After the agent finishes, run @im --verify or use the panel to check output against acceptance criteria.

agent operating layer

A prompt compiler and harness router, not a thesaurus.

One-Click Improve sits between your intent and the model. It normalizes a rough ask into the exact shape each harness can execute—with repo context, deliverables, acceptance criteria, validation steps, and guardrails. Agent-aware rewrites add the right discovery, patch discipline, and final-summary expectations without making you pick a harness manually.

surface awareness Routes for VS Code Chat, Cursor, Claude Code, Claude Cowork, Codex, and shell handoffs—with the right command syntax for each.
repo auto-context Infers stack, package scripts, generated folders, and agent instruction files when no .improv profile exists yet.
agent-aware rewrites Adds surface-specific discovery, patch discipline, validation commands, and final-summary expectations to every compiled spec.
task lifecycle Tracks each task from generated through applied, sent, and verified in .improv/tasks/.
outcome verification Checks agent final reports against the preflight task with @im --verify or the VS Code panel.
customizable skill Rewrite logic lives in canonical SKILL.md—read, edit, and sync the same skill across harnesses.
multi-harness sync Install writes the bundled skill to .claude/skills, .cursor/skills, .agents/skills, and .github/skills.
clarify over assume Optionally asks up to 5 clarifying questions when missing context would change the implementation—skippable when you need speed.

works where prompts start

Command routes for each environment

The same prompt-improvement engine is exposed through the shortcuts developers already reach for.

VS Code

Sidebar panel with side-by-side diff, inline edit, Apply to editor, Copy to chat, clarifying questions, post-agent verification, and @im in Chat. API keys use VS Code SecretStorage.

@im Fix login bug

Cursor

Slash commands for general and template-specific compilation—8 task variants.

/im Fix login bug /im-bugfix session expires on refresh

Claude Code

Global /im commands synced at install—same engine as Cursor and terminal.

/im Refactor dashboard loader

Claude Cowork

Full plugin parity: verify, tasks, templates, explain, and debug-route via cowork-im.

/one-click-improve:im Fix login bug

Terminal

Pipe specs into Claude Code, Codex, or CI. JSON schema, verify, and task memory built in.

@im "Refactor dashboard" --json

Codex

Copy or pipe compiled specs into Codex-ready workflows with surface-specific execution discipline.

@im "Add auth tests" | pbcopy

Harness Router

Ask which route and skill profile best fit the work before sending the task to an agent.

@harness "auth tests in Cursor"

Templates on the same engine: /im-bugfix /im-refactor /im-a11y /im-safe /im-fast /im-deep /im-test-heavy

before and after

The output is designed for execution

The improved prompt is not an essay. It is a work order with objective, context, constraints, deliverables, acceptance criteria, and validation steps.

Rough prompt

Fix login bug

Improved prompt


            

          

        

      


      

        

          

            
choose by task shape

            
Local for structure. Model-backed for ambiguity.

            

              local is the default: fast, offline, deterministic structure when the prompt is short or already names files and constraints.
              auto and model providers are for messy, domain-heavy asks where you want stronger scoping before the agent starts.
              VS Code keys live in SecretStorage; terminal keys stay 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 (readiness + weaknesses), compiled spec,
                token savings estimate, and memoryPath. Exit codes: 0 success, 1 error, 2 missing prompt.
              

              

                @im --json
                @im --tasks
                @im --verify
                @im --explain
                @im --debug-route
                @im --template bugfix
              

            

          

        

      


      

        

          

            
start here

            
Install once, improve prompts everywhere

            

              The installer detects VS Code and Cursor from your PATH or standard macOS app bundles,
              installs the extension, creates terminal shortcuts (@im, @harness, cowork-im),
              writes Cursor, Claude Code, and Cowork command files, syncs the bundled improve-prompt skill to
              .claude/skills, .cursor/skills, .agents/skills, and .github/skills,
              then optionally asks for a terminal provider key. You can skip it.
            

          

          

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