Fix with AI

Every blocker carries a paste-ready prompt for your AI build tool of choice. Restraint-engineered so the tool doesn't sweep-refactor or brick the codebase.

The shape of every prompt

PulseLight prompts follow a strict, restraint-engineered shape. The structure is identical across all eight supported tools so you build muscle memory:

<!-- pulselight prompt v2 -->

Goal:
<one-paragraph plain-English description of the issue>

Affected files:
- <path>:<line>

Working rules (enforced):
- <rule>
- <rule>

Do NOT:
- <thing the tool tends to do that breaks the codebase>
- <unrelated refactor>

Validation steps:
- <command to run>
- <expected outcome>

Output a unified diff. Summarise what changed in 3 bullets.

Why the “Do NOT” section matters

Most AI build tools, given a vague prompt, sweep-refactor — they reformat unrelated files, swap package managers, upgrade dependencies, rename variables. PulseLight prompts include an explicit anti-list calibrated to each issue. The result: the tool fixes the one thing you asked for, nothing else.

Supported tools

PulseLight generates the same prompt shape, with tool-specific framing tweaks, for:

  • Cursor
  • Claude Code
  • Lovable
  • Bolt
  • v0
  • Replit Agent
  • Windsurf
  • Aider

Plan mode (multi-issue)

For batches of related blockers (e.g. five Billable issues), PulseLight generates a plan-mode prompt: phase 1 fixes launch blockers, phase 2 fixes warnings, phase 3 cleanup. Tool-specific phasing keeps each step small enough that the AI tool can validate as it goes.

Telemetry & the loop

When you copy a prompt, PulseLight records that consumption against the finding fingerprint. If the same fingerprint disappears on the next scan, we mark the issue cleared via Fix-with-AI. That signal feeds the project’s Fix-progress card — the only KPI that matters is “did the prompt actually fix the thing?”