The cheapest place to fix AI slop is the plan

Engineers are getting buried in AI-generated PRs; most are full of slop.

The cheapest way to prevent slop is to give your agent a high-quality plan, before it starts writing code. Boris Cherny (the creator of Claude Code) recommends starting every complex task in plan mode:

Start every complex task in plan mode. Pour your energy into the plan so Claude can 1-shot the implementation.

Source

Getting a high-quality plan out of an agent requires iteration and feedback. This is what Boris means when he says “pour your engery into the plan”. The problem is it’s painful to give your agent feedback when all Claude gives you is a tiny TUI textbox at the bottom of the plan:

 ▐▛███▜▌   Claude Code v2.1.121
▝▜█████▛▘  Opus 4.7 (1M context) with xhigh effort · Claude Max
  ▘▘ ▝▝    ~/code/stripe-webhooks

# Plan: Idempotent Stripe webhook ingestion

## Context

Stripe occasionally double-applies `invoice.paid` events when retrying during a deploy. We need an idempotency layer that survives restarts and concurrent delivery.

## Steps

1. Add a `stripe_event_log` table keyed on `event.id` with `received_at`, `processed_at`, and `payload jsonb`.
2. On ingest, **insert the event row before signature verification** so we have a durable audit trail of every delivery attempt.
3. Dispatch the handler inside a transaction that stamps `processed_at`. If already set, short-circuit with 200.

## Handler sketch

```ts
export async function handleStripeWebhook(req, ctx) {
const raw = await req.text();
const event = JSON.parse(raw);
await ctx.db.insert(stripeEventLog).values({ id: event.id, payload: event });
const ok = ctx.stripe.webhooks.constructEvent(
  raw,
  req.headers.get('stripe-signature'),
  ctx.env.STRIPE_WEBHOOK_SECRET,
);
if (!ok) return new Response('bad sig', { status: 400 });
return dispatch(event, ctx);
}
```

Humans deserve better. We’re used to reading documents top to bottom and being able to quickly give feedback via inline comments as we read (ala Google Docs). That’s why we built PlanBridge.

Provide precision feedback with PlanBridge

PlanBridge makes it easy to highlight a specific word, line, variable name, or entire section of the plan and provide precise feedback to the agent. It sends your detailed feedback back to the agent to refine the plan before any code is generated. It works alongside your existing workflow and coding harness, with no desktop app or heavy workflow to follow.

How it works

1. Install PlanBridge on your machine. Our installer automatically hooks into your coding harness ¹.
2. Enter plan mode and type your prompt. When the agent produces a plan, PlanBridge automatically loads in your browser.
3. You add precise feedback on improvements to the plan.
4. Your feedback is sent back to the agent for review. Once you’re happy, the agent starts coding.

Fixing bad code during code review is expensive

AI generates code faster than humans can review it.

• PR volume per developer is up 98% in high-AI-adoption orgs. (Faros AI, 2025)
• PR review time is up 441%, with 31% more PRs merging with no review at all. (Faros AI, 2026)
• 66% of developers say AI code is “almost right but not quite.” (Stack Overflow, 2025)

We see that PR volume continues to climb while bandwidth to conduct quality code review remains the same. Catching issues at plan time causes agents to generate higher-quality code the first time.

Fix bad plans, not bad PRs

PlanBridge is one of the highest ROI tools for AI that we’ve used. It’s free and open source (soon!).

Want to help build the future of PlanBridge? Please join our Slack Community or get in touch via email.