4-week build Capstone Track

Ship something real.

Four weeks. One project. Every practice in the catalog points at this moment. The one where you stop reading about AI and ship something that uses it. Pick a target, follow the path, hold yourself to the Friday checkpoints.

Zero new content; curation only. This is the ordered tour through the practices that gets you from "I'm interested" to "here's the URL of what I shipped." If you only do one track on Pradhya, do this one.

Pick the target you'll ship.

Pick exactly one. Be specific. The track is the same shape regardless. Only the project changes.

Track A · Operator

A Claude system for your week

The Monday Setup, expanded into a personal AI cowork system. Start with a one-week time baseline, then run the system for two weeks and measure what actually changed.

Done means: two complete weekly runs, a before/after time log, and one revision based on a failure you observed.

Track B · Builder

An agent that ships a task

An autonomous Claude agent that does one real job end-to-end. Eval harness, observability, real prompts, deployed somewhere your colleagues can use it.

Done means: a versioned eval set, a passing safety gate, and five task runs reviewed by at least two intended users.

Track C · Founder

A working SaaS prototype

Landing page, sign-up, a clearly priced offer, and one AI feature that delivers the core value. Payment integration is optional; if included, keep it in test mode during the capstone.

Done means: five target users complete the core flow, you record where they stop, and you improve the most common failure.

Track D · Tinkerer

A second brain on a Pi

A personal AI assistant running on hardware you own. It works from an approved test folder, summarizes your week, and clearly labels which features send data to a cloud model.

Done means: three successful runs on the device, a documented recovery test, and a privacy review of every external connection.

Week 1 · Orient

Foundations & first artifact.

Get the mental model + the working Claude setup. Produce your first artifact by Friday: a Claude Project that you'll use again.

Day 1Read the free Monday Setup. Build the Claude Project. Run it Sunday night.
Day 2Practice 11 · Prompt Engineering Units 1-4. Save 3 prompts to a personal library.
Day 3Practice 17 · Under the Hood. Build the mental model. Print the cheatsheet.
Day 4Practice 03 · Cowork Units 1-3. Connect at least 3 connectors. Run R01.
Day 5Pick your target (A/B/C/D above). Write one paragraph: by Friday week 4, the user can ___.
Sat-SunUse what you've built. Time-box: 2 hours. Note what worked, what felt clunky.

Ship by Friday: a personal Claude Project (for Track A) OR a written spec of what you'll build (for B/C/D). Send to one person for feedback.

Week 2 · Build

The first real version.

Write the first version of what you'll ship. Crude. Working. Not pretty. By Friday, you can demo a 30-second version of the thing.

Day 1Practice 02 · Claude Code Units 1-4. Install Claude Code. Make your first commit using it.
Day 2Track-specific deep dive:
A → Practice 03 (rest of Cowork) + 13 (Memory)
B → Practice 04 (Agents) Units 1-6
C → Practice 15 (Real SaaS) Units 1-4
D → Practice 05 (NanoClaw). Set up the Pi
Day 3First-draft version. Crude is fine. Wire it together. Don't polish.
Day 4Run it end-to-end. Note what breaks.
Day 5Fix the top 3 breaks. Stop. The fourth break is week 3 territory.
Sat-SunUse it yourself with synthetic or non-sensitive data. Record every failure. Do not invite outside users until the Week 3 safety gate passes.

Ship by Friday: a demo-able first version. Recordable in 30 seconds. Imperfect. Self-tested end-to-end with failures captured for Week 3.

Week 3 · Harden

Make it shippable.

Move from "works for me" to "works for a stranger." Real prompts. Real validators. Real failure handling.

Day 1Practice 19 · Safety in Production. Run the red-team prompt against your real system prompt, document data boundaries, and fix every release-blocking issue.
Day 2Practice 12 · Context Engineering. Apply the audit to your project. Reduce irrelevant context without removing evidence the task needs.
Day 3Practice 14 · Observability. Add the LLM-as-judge eval. Make sure you can answer "what did it do for user X yesterday?"
Day 4Practice 18 · Multi-Agent Systems, but only if your evals show the project genuinely needs it. Otherwise, keep the simpler design.
Day 5Practice 20 · The Agent Workbench. Install the Workbench Pack, then run the release gate: safety checks, task evals, recovery path, and a manual review. Invite a pilot only if every blocking check passes.
Sat-SunPilot with 3 friends. Capture every failure they hit. They're more valuable than a hundred internal users.

Ship by Friday: a hardened version. Passes the 8-question safety checklist. Has at least one eval that runs on a schedule.

Week 4 · Launch

Send the link.

Stop adding features. Start collecting evidence. Public tracks need target-user sessions; personal tracks need repeated use plus a trusted walkthrough.

Day 1For Track C / B: deploy to a real URL. For A / D: write the README that explains how to use it.
Day 2Practice 16 · Founder's Playbook Units 1-3 (even if you're not a founder; it teaches launch-week discipline). Write the launch checklist.
Day 3Tracks B/C: send the link to 5 specific target users. Tracks A/D: complete two real-use sessions and walk one trusted reviewer through the system without exposing private data. Ask: "Tell me one thing that was confusing and one thing that worked."
Day 4Read every reply. Don't defend. Note patterns.
Day 5Pick ONE thing from the feedback. Fix it. Send back a note: "Here's what I changed based on what you said."
Sat-SunDecide: keep building, hand off, or kill. All three are honest answers.

Ship by Friday: meet the "done means" evidence for your track, capture written observations, make at least one evidence-based improvement, and decide whether to continue.