Stop Describing Errors. Paste Them Into the Agent.
Retyping a bug in your own words throws away the best evidence. The paste-the-error workflow: what to copy, what context to add, when to push back.
Watch how experienced developers work with AI agents and one habit stands out: they paste. Non-programmers tend to narrate — "the login button seems broken, maybe the form isn't sending?" — while engineers dump the raw error and its stack trace into the chat. The engineer's way wins because error output is dense evidence: exact file names, line numbers, the failing call chain, the precise complaint. Narration is a lossy translation of that evidence through someone who doesn't yet know which details matter. Your job isn't to interpret the error. It's to deliver it intact.
Copy the whole thing, especially the ugly parts
Grab the complete output, not just the last line. A stack trace reads top-down: the first frames name what immediately failed, and the frames beneath show how execution got there — trimming it deletes the story. Keep the alarming red text people instinctively delete; it's usually load-bearing. Include any warnings printed before the error, too, since the actual cause often shows up earlier than the crash does.
Alongside the error, capture what you did right before it appeared: the exact click, the URL, the values you entered. Three concrete reproduction steps beat any amount of theory about why it broke. If nothing visible showed up on screen, open the browser's developer tools — the console tab collects JavaScript errors, and the network tab lists requests that failed along with their status codes. Both are pasteable evidence.
Connect the Claude or Codex you already pay for — the rest runs on workers that cost a fraction.
Download meshcode →Add two sentences of context, not a paragraph of theory
Sometimes the error alone is ambiguous, and short factual context helps enormously: what you expected versus what happened, whether this ever worked before, and what changed recently. "This started after you renamed the checkout component" is close to gold — recent changes are the leading suspect in most bugs. What helps less is extended diagnosis: "I think the database might be timing out because of the new feature?" Speculative theories can anchor the agent onto the wrong track, and interpretation happens to be the part it's genuinely better at than you, for now. Facts in, theories out.
Ask the agent to reproduce before it fixes
The strongest pattern is politely refusing instant fixes. An agent that answers a pasted stack trace with a plausible patch — without running anything — is frequently solving a different bug than yours. Ask it to reproduce first: run the failing command, hit the endpoint, execute the test, and confirm the same error appears before any code changes. Agents with filesystem and terminal access can do much of this themselves, reading logs, running the app, and probing endpoints directly, provided you've set that access up sensibly; there's a checklist for granting a coding agent filesystem access without over-exposing your machine. Reproduction converts debugging from guessing into verifying.
Judge the fix, then re-test with fresh evidence
When a fix lands, do two things. First, rerun your original reproduction steps and confirm the bug is gone, then spend thirty seconds clicking around neighboring features, because hurried patches sometimes break something adjacent. Second, commit the working state so there's a save point before the next experiment. If the same problem resurfaces, don't append fragments of the old output — paste the complete new error, since it may differ from the first in ways that matter. And before shipping a meaningful fix toward real users, a structured pass of reviewing AI-generated code against what changed is cheap insurance.
The meshcode angle
Debugging is a copy-and-paste loop, which makes workspace layout matter more than people expect. In meshcode, a native desktop app for macOS and Windows, every pane runs its own agent session over the same repository: paste the stack trace into one pane while another independently investigates the same failure, and compare which hypothesis holds before anyone touches code. Drive it with the Claude Code or Codex CLI subscription you already have, or start with meshcode's own metered models, billed pay-as-you-go with no monthly fee.
👉 Download meshcode — Mac, Windows