The Night My Local LLM Became My Project Manager (and Actually Delivered)

July 15, 2026

It started as a harmless experiment at 11:47 PM: "Could my local LLM help me plan tomorrow?" I was behind on a small side project-a CLI tool I'd promised a friend-and my brain was doing that special midnight thing where everything feels urgent and impossible.

I run a local model for privacy and speed (no API calls, no "rate limit exceeded" drama). I opened my terminal, fed it the mess in my head-half-finished tasks, vague goals, and one hard deadline-and asked: "Act as my project manager. Make a plan that I can follow in the morning."

Within two minutes, it was assigning me work like it had a clipboard and a grudge.

The Setup: Turning a Chatbot into a PM

The trick wasn't "be smarter," it was "be specific." I gave it three inputs:

1) The goal: "Ship v0.1 of the CLI tool by Friday. It must: parse a config file, run two commands, and output a report."

2) Constraints: "I have 90 minutes tomorrow morning, 45 minutes at lunch, and 2 hours at night."

3) Current state: "Parsing works. Command runner is flaky. No tests. README is blank."

Then I asked for a plan in a format I could execute: "Return a checklist, timeboxed, with dependencies and a definition of done for each task."

What I got back looked like an actual sprint plan: a staged checklist (Morning/Lunch/Night), each item with acceptance criteria, and a warning about risk ("command runner flakiness will cascade; fix first"). It also suggested a tiny scope cut: "Defer fancy formatting in reports; output JSON first." That one decision saved my Thursday.

The Night It Started Managing Me

Here's where it got weirdly effective: I kept it open while I worked, and it became a live PM.

When I said, "The command runner fails on Windows paths," it didn't just sympathize-it asked a clarifying question: "Are you shelling out through a single string or argv array?" Then it recommended a concrete change and a mini-test matrix: "Test with spaces, unicode, and quoted args."

When I drifted into polishing the README at 1:30 AM, it called me out: "This is not on the critical path. Move README to 'Night block' after tests." Rude. Correct.

The most useful pattern was a tiny loop:

I paste what changed (a diff snippet or a short description)
It updates the plan (what's done, what's blocked, what's next)
It writes the next prompt (a command to run, a test case, or a commit message)

Example: I asked for "a commit message that matches conventional commits," and it generated: `fix(runner): handle paths with spaces by using argv array`.

By 2:13 AM, my backlog had turned into a clean kanban: "Now / Next / Later," with clear definitions of done. I didn't feel like I was "chatting with AI." I felt like I had borrowed someone's executive function.

Practical Tips If You Want Your Own LLM-PM

If you try this, make it earn its keep:

Force structure: Ask for outputs in tables or checklists with time estimates and dependencies.
Insist on acceptance criteria: "Done means X passes and Y is documented."
Add a status ritual: Every hour, paste: "What I finished / what's blocking / what I'm about to do."
Use scope cuts deliberately: Ask: "What can I defer without breaking v0.1?"
Keep it local for sensitive work: Specs, customer notes, and logs stay on your machine.

The next morning, I didn't wake up to a guilt fog. I woke up to a plan that was already written-and a tiny, tireless project manager waiting in my terminal, ready to ask the question I usually avoid: "What's the next smallest thing?"

Search This Blog

tylers-blogger-blog