diff --git a/packages/framework/FEATURES.md b/packages/framework/FEATURES.md
deleted file mode 100644
index 442a243..0000000
--- a/packages/framework/FEATURES.md
+++ /dev/null
@@ -1,32 +0,0 @@
-# Features
-
-> [!NOTE]
-> **Potential features, for evaluating the USP potential** of The Framework. This **isn't a roadmap**, it's just brainstorming to shape a vision for The Framework.
-
-- Autopilot: minimum intervention, automatic code refactoring (for maintainability), automatic security audit, automatic new feature requests (related: product manager agent), automatic feature request implementation
-  => potentially groundbreaking USP for non-engineers (maybe even engineers)
-- Technical control => review only
-  - Highly polished PRs:
-    - Nice three-level PR overview: TLDR, summary, and details
-    - Minimal changes in PR (clean refactoring in susbsequent PR)
-  => nice USP (potentially significant USP)
-- Bunch of `.md` files (`DECISIONS.md`, `ARCHITECTURE.md`, `BRAINSTORMING.md`, `KNOWLEDGE_BASE.md`, ...) as seam between AI and humans
-  => massive USP
-- Product manager agent
-  => ?? No clue how much USP, but potentially massive USP — to be tested. (Because AI is bad at writing documentation, I suspect it's going to be bad at being a product manager. But let's see.)
-- High-quality prompts, for example:
-  - [Highly-effective code refactoring prompt](https://gist.github.com/brillout/8abfd310bad5df422ae56c5c9066ffc5)
-  - [Anti-laziness boost](https://gist.github.com/brillout/8cf6b2b3b4708d0a4cac192b87cfb97c)
-  => significant USP (potentially massive?)
-- Auto triggers
-  - CI red => trigger agent
-  - Sentry red => trigger agent
-  => significant USP
-- Cron jobs to max-out daily usage limits
-  => significant USP
-- Bootstrap mode
-  => nice USP (potentially groundbreaking, if we manage to make AI autonomously create advanced apps)
-- Notifications
-  => nice USP
-- Mobile app
-  => nice USP
diff --git a/packages/framework/PLAN.md b/packages/framework/PLAN.md
new file mode 100644
index 0000000..07bd4dd
--- /dev/null
+++ b/packages/framework/PLAN.md
@@ -0,0 +1,94 @@
+# Roadmap 🚀
+
+> [!NOTE]
+> - Also includes brainstorming ideas (listed at [Untriaged](#untriaged) and [Low prio](#low-prio)).
+
+## High prio
+
+Selected: implement now-ish.
+
+- Usage limit maxing
+  - VALUE-HIGH
+  - Queues: global queues of prompts to be fired whenever there's capacity
+  - Cron jobs to max-out daily usage limits (fires queues + maintenance prompts)
+  - Implementation
+    - Accessing usage limits: https://github.com/gemstack-land/gemstack/pull/300#issuecomment-4918256151
+  - Marketing
+    - Very easy sell (very sexy feature, massive added value)
+    - Unique USP (since Claude has no interest to implement this)
+    - Attracts very prolific AI users (more likely to be early adopters of The Framework)
+    - Landing page illustration idea: progress bar "used AI daily budget" (without The Framework => red bar, with The Framework => green bar, a "candy" progress bar that looks like a skills progress bar in MMORPG video games)
+  - Ideas/brainstorming (don't implement yet)
+    - `Auto`-model chooser (use cheap model to analyze prompt and select the right model)
+    - Allow user to select max quota budget (in %) per prompt
+- Don't reapeat yourself
+  - VALUE-HIGH
+  - AI-human seam files: `KNOWLEDGE_BASE.md`, `DECISIONS.md`
+    - TODO: maybe develop system prompts to ensure these are maintained?
+- Dashboard
+
+## Medium prio
+
+Candidates: implement soon? Maybe even now-ish?
+
+- Anti-lazy-pill
+  - VALUE-HIGH
+  - https://github.com/gemstack-land/gemstack/issues/297#issuecomment-4913683778 (the `PLAN.md` + `TODO.md` trick)
+  - TODO: install it as system prompt injected to *all* prompts (let's see if it's somtimes counterproductive)
+- Bootstrap mode
+  - VALUE-HIGH (potentially groundbreaking, if we manage to make AI autonomously create advanced apps)
+- Auto triggers
+  - VALUE-HIGH
+  - GitHub CI red => trigger agent
+  - Sentry red => trigger agent
+  - TODO: check if hard to implement?
+  - Implementation
+    - Bad: requires integrations with many providers (not every company uses GitHub)
+
+## Low prio
+
+Postponed: don't implement yet.
+
+- Mobile app
+  - VALUE-LOW
+- Scaling
+  - For large codebases: `CODEBASE_OVERVIEW.md`
+  - VALUE-LOW
+
+## To research
+
+- Product manager agent
+  - VALUE-?? (no clue how much value, but potentially massive value — to be tested. Because AI is bad at writing documentation, I suspect it's going to be bad at being a product manager, but let's see.)
+  - Bunch of `.md` files as seam between AI and humans? `BRAINSTORMING.md`, `FEATURE_REQUEST.md`
+
+## Untriaged
+
+TODO: sort the following in high/medium/low prio
+
+- Autopilot mode: minimum intervention, automatic code refactoring (for maintainability), automatic security audit, automatic new feature requests (related: product manager agent), automatic feature request implementation
+  - VALUE-HIGH (potentially groundbreaking)
+- Technical control mode => review only
+  - VALUE-MEDIUM
+  - Highly polished PRs:
+    - Nice three-level PR overview: TLDR, summary, and details
+    - Minimal changes in PR (clean refactoring in susbsequent PR)
+- Maintenance
+  - VALUE-MEDIUM
+  - High-quality prompts:
+    - [Highly-effective code refactoring prompt](https://gist.github.com/brillout/8abfd310bad5df422ae56c5c9066ffc5)
+      - Let's break this prompt in two: one for maintainability (e.g. DRY), and a second one for readability (so that humans can easily read the code)
+        - Try the readability prompt on brand-the-framework (it has lots of potential for top-down code structure refactoring)
+    - Security audit (TODO: develop scurity audit prompt)
+- TLDR thinking out loud
+  - VALUE-MEDIUM
+  - Show TLDR of the model's thinking (=> nice overview of all the thinking done during this session)
+  - Also show live thinking (same thinking-out-loud as Claude Code => just forward the Claude Code CLI output)
+- Notifications
+  - VALUE-MEDIUM
+- Give AI access to GitHub issues/discussions
+  - VALUE-HIGH
+  - Implementation
+   - Ideally, AI can access the entire text of all issues (and optionally even all comments)
+   - Complex to implement?
+  - Alternative: don't use GitHub issues at all for product management?
+  - Maybe sync between GitHub issues and a repo file.