Getting feedback into Claude Code

Claude Code reads files in your repo, so the cleanest workflow is to drop a feedback.md file into the project root and reference it in your prompt.

This page is part of the guides hub on connecting visual feedback to AI coding agents. It is the Claude Code-specific playbook; the cross-agent version lives at how to give feedback to an AI coding agent.

What you'll learn

  • The three ways to get feedback into Claude Code
  • Why the file-in-repo approach wins for most workflows
  • A concrete walkthrough from staging issue to applied fix
  • How the .md persists across sessions and commits

The three ways to get feedback into Claude Code

1. curl the markdown into your repo. The cleanest path. CobaltCapture's review URL has a /markdown endpoint that returns the raw markdown. One command lands the file in your repo root:

curl -fsSL https://cobaltcapture.com/r/<slug>/markdown > feedback.md

Now feedback.md is a first-class repo artifact. Claude Code reads it like any other source file.

2. Paste raw markdown into the prompt. Open the review in your browser, copy as markdown, paste into Claude Code. Works for a single session but the content is session-scoped, next session starts from scratch.

3. Reference the URL directly in the prompt. Claude Code can fetch URLs, but the file-in-repo approach is canonical because it persists. A URL reference works for one-shot but is harder to re-find tomorrow.

For ongoing work, option 1 is the right default. The other two are escape hatches for when you do not want the file in the repo.

Why .md in repo wins

Claude Code's mental model is file-aware. It reads package.json to understand your stack, README.md to understand the project, src/components/Button.tsx to understand the code. A feedback.md in the project root fits naturally into that model. You can prompt "Read feedback.md and address items in order" and the agent does exactly that.

Three properties make this the canonical workflow:

The file persists across sessions. Open Claude Code tomorrow, reference the same file, get the same context. URL-based approaches do not survive compaction or new sessions.

It commits with the work. Each finding has an audit trail tied to a real commit. Three months later when someone asks "why did we change the submit button radius," the commit message references feedback.md and feedback.md still has the screenshot and your dictated rationale.

It survives context-window compaction. Claude Code aggressively compacts. A pasted block can drop out; a file reference does not, the agent re-reads the file when it needs to.

A walkthrough

You are in your project on the terminal. Open your staging build in a browser. Open CobaltCapture in another tab. Capture findings: drag a box around each broken region, dictate the context, repeat. Hit Publish.

You get a URL: cobaltcapture.com/r/abc12345. In your terminal at the repo root:

curl -fsSL https://cobaltcapture.com/r/abc12345/markdown > feedback.md

The file lands. Open Claude Code:

Read feedback.md and address items 1 through 5 in order. For each
item: summarize what you understood, propose the fix, apply it, and
run any related tests. Pause for confirmation between items.

Claude Code reads feedback.md as a first-class file. It treats each H2 section as a discrete finding, follows the embedded image URLs when it needs visual context, and works through the document in order. The dictated context in each section, the why, the viewport, the browser, becomes part of how the agent reasons about the fix.

When the work is done, commit feedback.md with the changes. The PR description references it. The PR review has the same artifact your reviewer captured. Three months later it is still there.

How CobaltCapture fits in

CobaltCapture publishes the review and exposes the markdown endpoint that the curl command depends on. The capture flow takes about a minute for three to five findings, voice dictation handles the context without typing, and the resulting .md is structured for exactly this workflow. See feedback for Claude Code for the dedicated workflow page.

Frequently asked questions

How do I get a CobaltCapture review into Claude Code?

Use curl to download the markdown export of the review into your repo root, then reference the file from your Claude Code prompt. The agent treats it as any other repo file.

Does Claude Code follow URLs?

Claude Code is file-first and works best when feedback lives as a file in the repo. URL-following exists but the file-in-repo approach is canonical because it survives session compaction and gets committed with your work.

Can I just paste markdown into the terminal?

Yes, but it is less clean than saving to a file. Pasted markdown in the terminal is session-scoped. A feedback.md file in the repo persists, gets committed, and is referenceable across future sessions.

Where should the feedback.md file live in my repo?

Project root is the canonical location. Claude Code finds it without help and the file shows up cleanly in commits. Subdirectories like docs/feedback/ work too if you want a record of multiple reviews over time.

Will the feedback.md file survive session compaction?

Yes. The file lives in the repo, not the session. Even if Claude Code compacts the context window, you can re-reference feedback.md in your next prompt and the file is still there.

Frequently asked questions

How do I get a CobaltCapture review into Claude Code?

Use curl to download the markdown export of the review into your repo root, then reference the file from your Claude Code prompt. The agent treats it as any other repo file.

Does Claude Code follow URLs?

Claude Code is file-first and works best when feedback lives as a file in the repo. URL-following exists but the file-in-repo approach is canonical because it survives session compaction and gets committed with your work.

Can I just paste markdown into the terminal?

Yes, but it is less clean than saving to a file. Pasted markdown in the terminal is session-scoped. A feedback.md file in the repo persists, gets committed, and is referenceable across future sessions.

Where should the feedback.md file live in my repo?

Project root is the canonical location. Claude Code finds it without help and the file shows up cleanly in commits. Subdirectories like docs/feedback/ work too if you want a record of multiple reviews over time.

Will the feedback.md file survive session compaction?

Yes. The file lives in the repo, not the session. Even if Claude Code compacts the context window, you can re-reference feedback.md in your next prompt and the file is still there.

Capture your first review.

About a minute from open tab to a shareable URL your agent can ingest.

Start capturing