The Blueprint Before The Build

Nexus Phase 0: planning that earned its keep

The Frame

The win was real — but locked behind a terminal.

Amplifier proved AI-augmented work is real. But its power lived behind a developer CLI, reachable only by the terminal-fluent. Reaching everyone else — PMs, analysts, communicators — meant a different kind of thing: a desktop Product.

So how do you start building something you're not yet sure how to build?

The Proof

It opened with a blueprint, not a splash of code.

On 2026-01-28 the amplifier-nexus repo's initial commit (e1485c1) landed ~26,331 lines — ~15,485 of them Markdown. The bulk of the opening work was written planning.

But what, exactly, was being planned?

26,331 lines

Total lines in the initial commit

15,485 md

Of those, Markdown lines

The Project

Nexus is Amplifier grown into a desktop AI teammate.

Its README frames it as "the evolution of Amplifier from developer CLI to full AI teammate" — one that "lives in your workspace (not the browser, not an IDE)." Nexus is the vehicle for reaching non-terminal users.

That vehicle sat at a fork with no obvious answer.

The Tension

Two shapes of product — and picking wrong burns months.

A terminal-centric Distro (amplifier-distro) versus a desktop Product for knowledge workers. Both had parallel Phase 0 efforts in early 2026. Commit code to the wrong shape and you burn months of implementation.

The answer: decide on paper before deciding in code.

The Distro

amplifier-distro: makes Amplifier "just work" across CLI, TUI, and Voice — but still needs Python, Node, git, and gh CLI.

The Product

Nexus: a desktop app for knowledge workers, living in your workspace — no terminal fluency required.

The Mechanism

Phase 0's first move was 12 planning docs, 7,264 lines.

The blueprint spanned vision, positioning, GTM, metrics, roadmap, requirements, and risks — plus a 915-line PHASE-0-FOUNDATION spec defining the desktop-app foundation on amplifier-core and amplifier-foundation.

Deciding what to build, and for whom, came first.

12 docs / 7,264 lines — formal product & roadmap planning
Coverage — vision, positioning, GTM, metrics, roadmap, requirements, risks
915 lines — PHASE-0-FOUNDATION spec on core + foundation

The Turn

Planning-led, not planning-only — code came next day.

On 2026-01-29 commit b35f958 landed a working Next.js scaffold: 20 TypeScript files, 2,525 lines, co-authored with Claude Sonnet 4.5. Yet the center of gravity stayed on deciding, not grinding.

The plan was real enough to be tested against reality.

# 2026-01-29  commit b35f958
"Phase 0 implementation.
 Moving on to Phase 0 testing."

# scaffold delivered
src/  20 TS files, 2,525 lines
Co-Authored-By: Claude Sonnet 4.5

The Payoff

The plan proved its worth by being thrown away.

Weeks later, on 2026-03-04, the first design was declared "obsolete ... previous design exercises." Commit 7d41764 deleted 42 files (~29,944 lines). The next day, 34aed71 added redesign specs (specs/00–07, +4,810 lines).

A blueprint judgeable enough to discard made the "not yet" cheap.

42 files

Deleted 2026-03-04 (7d41764)

29,944 lines

Removed as obsolete design

+4,810 lines

Redesign specs, next day (34aed71)

The Takeaway

Plan hard enough to be wrong — cheaply.

A deliverable that ships nothing but a judgeable blueprint can be worth more than the code it defers. The honest "not yet," captured solidly enough to be judged and discarded, spared the team a costly wrong build.

The blueprint's job was to be throwable. It was.

Sources

Research Methodology

Data as of: verified against the marklicata/amplifier-nexus repo (local clone; origin git@github.com:marklicata/amplifier-nexus).

Phase 0 status: blueprint DELIVERED (12 docs / 7,264 lines + 915-line foundation spec); code scaffold DELIVERED but provisional; first design later declared obsolete and superseded.

Research performed (git / gh / grep):

First commit + repo creation: git log --all --reverse --date=iso --pretty=format:'%h|%ad|%an|%s' and gh repo view marklicata/amplifier-nexus → e1485c1, 2026-01-28.
Line totals of initial commit: git show e1485c1 --numstat | awk '{added+=$1} /\.md$/{md+=$1} END{print added, md}' → 26,331 total / 15,485 Markdown.
Planning-doc count: git show e1485c1 --numstat | grep -E 'docs/(product-planning|roadmap-planning)/.*\.md' → 12 files, 7,264 lines.
Foundation spec size: git show 015cb58:.amplifier-app-docs/02-roadmap-planning/archive/PHASE-0-FOUNDATION.md | wc -l → 915 lines.
Code scaffold: git show b35f958 --numstat | grep -E 'src/.*\.(ts|tsx)$' → 20 TS files, 2,525 lines; co-author trailer "Claude Sonnet 4.5".
Discard + redesign: git show 7d41764 --numstat; git show 34aed71 --numstat → 42 files / 29,944 lines removed (2026-03-04); specs/00–07 +4,810 lines (2026-03-05).
Distro contrast: grep -niE 'CLI, TUI, Voice|Python, Node, git' OPINIONS.md README.md in amplifier-distro.

Gaps: The "~9–13 days" of planning-led work is derived from commit timestamps (Jan 28 → Feb 6 = 9 days; → Feb 10 = 13 days), not a repo string. "Distro-vs-Product fork" and "months saved" are narrative framing, not verbatim repo text.

Primary contributors: Mark Licata (project lead, 56+41 commits); Alex Lopez (MVP planning data, 1 commit); Claude Sonnet 4.5 (co-author of implementation commit b35f958).