Let Claude Code Decide What to Build Next

If you build with an AI agent, you’ve probably noticed the ceiling: your agent is spectacular at building and useless at choosing. It can implement anything you describe in an afternoon — but ask it “what should I work on this week?” and it has nothing, because everything it knows about your product comes from the code. The users, the requests, the votes, the frustrated comment threads — all invisible to it.

This guide fixes that. By the end, your agent — Claude Code, Claude Desktop, Codex CLI, Cursor, or any MCP-compatible client — will be able to read your feedback board directly: pull the most-voted requests, read the discussion behind them, weigh demand against implementation effort in your actual codebase, and even update statuses when you ship. Prioritization stops being a gut-feel afternoon and becomes a conversation with receipts.

The prerequisite: feedback that lives in one place

None of this works if your feature requests are scattered across DMs, support email, and memory. The foundation is a feedback board — one front door where requests land, users vote, and duplicates merge, so the vote counts actually mean something. If you don’t have that yet, start there (here’s the full system); the agent workflows below are only as good as the signal underneath them.

The second prerequisite is that your board speaks MCP — the Model Context Protocol, the open standard AI assistants use to call external tools. Most serious feedback tools ship an MCP server now. The examples below use FeatureWish’s (seven tools, one bearer token, part of the $20/mo Pro plan), but the workflow translates to any board with an MCP endpoint.

Step 1: Connect the board to your agent

With FeatureWish, generate a token in Settings → MCP, then register the server. In Claude Code, that’s one command:

claude mcp add --transport http featurewish   https://app.featurewish.com/mcp   --header "Authorization: Bearer fw_mcp_YOUR_TOKEN"

In Cursor, add the same URL and header to .cursor/mcp.json; in Codex CLI, add it to ~/.codex/config.toml under [mcp_servers.featurewish]; in Claude Desktop or Claude.ai, add it as a custom connector. Your agent discovers the available tools by itself on first connect — for FeatureWish that’s list_boards, list_posts, get_post, get_roadmap, get_changelog, create_post, and update_post_status. No SDK, no glue code.

Sanity-check the connection with something trivial:

“List my feedback boards and tell me how many posts are on each.”

If you get an answer, you’re wired up.

Step 2: Ask the question you actually care about

Here’s the prompt that makes the whole setup worth it:

“Look at the ten most-voted open requests on my feedback board. For each one, skim the relevant parts of this codebase and estimate rough implementation effort — small, medium, large. Then rank them by votes-to-effort and tell me which three you’d do first, with reasoning.”

Read what’s happening there: the agent combines demand data it pulls from the board with effort estimates it derives from your actual code — a cross-reference no feedback dashboard can do, because no dashboard can see your codebase. The output is a ranked shortlist you can interrogate (“why do you think SSO is a large?”) rather than a static chart.

Two honest caveats. First, votes are a starting signal, not a mandate — they overweight the loud and underweight your highest-value customers, so tell your agent about anything it can’t see (“weight requests mentioned by paying customers higher; the enterprise trial asked for SSO twice”). Second, effort estimates from a code-skim are rough. Treat the ranking as a draft to react to, not a verdict — reacting to a concrete draft is still ten times faster than starting from a blank backlog.

Step 3: Turn the winning request into a spec

The comments under a popular request are the best spec material you own: real users describing real situations, in their own vocabulary, with edge cases you’d never invent. Mine them:

“Get the ‘CSV export’ post with its full comment thread. Draft an implementation plan: what people actually need (quote them), what’s explicitly out of scope, and acceptance criteria. Flag anything the comments disagree on.”

The agent pulls the post and every comment, and the spec it drafts is grounded in what fourteen real people said — not in what you remember them saying. Because it’s the same agent that will write the code, the spec flows straight into implementation in the next message.

Step 4: Close the loop from the same chat

The loop is what makes feedback compound: users see that asking works, so they keep asking. Its weak point has always been the last step — you ship, then forget to tell anyone. With the board connected, closing the loop is part of shipping:

“We just merged CSV export. Mark the request completed on the board and draft a changelog entry from the PR description.”

On FeatureWish, flipping a request to Completed automatically emails everyone who voted for it. That’s the trust-building moment most products fumble, executed in the same window where you merged the code, seconds after merging it.

Step 5: Make triage a standing prompt

The board stays trustworthy only if someone keeps it tidy. That someone can now be the agent, on a schedule or whenever you sit down:

“Check my feedback board. Summarize anything new since Monday, flag likely duplicates of existing requests, and tell me if any planned item is picking up unusual vote velocity.”

Five minutes of reading a digest replaces twenty minutes of clicking through posts — and a tidy board makes every other prompt in this guide smarter.

What to keep for yourself

A note on judgment, because “let your agent decide” is deliberately provocative: the agent reads, ranks, drafts, and files. The decisions stay yours. It doesn’t know your strategy, your runway, or that the quiet customer who never votes pays a third of your MRR. The right mental model is a chief-of-staff who arrives at every planning session having already read everything — not an autopilot.

That’s also why we’d be wary of tools where an AI silently reprioritizes your roadmap for you. The value isn’t automation of the decision; it’s the collapse of the distance between the demand signal and the person deciding.

The setup, recapped

1. Get feedback flowing into one board with voting (free on FeatureWish, forever).
2. Turn on the MCP server (Pro, $20/mo flat) and generate a token in Settings → MCP.
3. Register the endpoint with Claude Code, Claude Desktop, Codex CLI, or Cursor — copy-paste snippets here.
4. Steal the prompts above, starting with the votes-to-effort ranking.

For the wider picture — why agent-readable is becoming table stakes for founder tools, and the full set of workflows — see FeatureWish for AI agents.