Build record — Erick Manrique

One page · every flagship opens its own dossier

01Projects

What I built.

Two flagships and five supporting systems. Each carries an honest status chip, one load-bearing number or a statement in its place, and the stack it actually runs on.

REC-01Flagship

Ambito

A permission-aware context layer that turns an organization’s communications into structured, persistent memory — served to both the people and the agents doing the work.

Postgres + pgvector
MCP server
Haiku · Sonnet · Opus
Gmail · GCal · Zoom

Deployed prototype · staging · no external users

No external users, no revenue. Production promotion has not been run.

FIG 1.11,145

backend tests passing, plus 120 end-to-end

Three-model pipeline; agents read only what the requesting person can read.

REC-02Flagship

Agentic build system

A custom TypeScript MCP orchestrator. The agents are stateless — they coordinate only through a shared handoff file, so the build moves from spec to commit without me pasting between tools.

TypeScript
Claude Agent SDK
MCP
Swift
Self-hosted CI

Paused

No shipped app, no users — the engineering process is the artifact, not traction.

FIG 1.2Commit-on-green gate

A commit lands only if the test run exits 0 — a rule I added after a build-only gate let three failing-test commits reach main.

Supporting systems5 records

REC-03Supporting

CoBrain

A minimal-loader context runtime for AI sessions: a two-file boot, a generated corpus map, and a fetch router keyed on trigger vocabulary. It runs under every project on this page.

Claude plugin
Manifest builder
Write-leases

Personal runtime · 3 repos onboarded

−52%resident context on the Portfolio onboarding — nothing lost, everything one trigger away

A personal system, not a product — no external users.

See section 02

REC-04Supporting

StudyAgent

A Slack-only spaced-repetition tutor: SM-2 mastery tracking, Claude-generated questions, and a confidence rating collected before the answer is revealed.

Node
Slack Bolt
BullMQ + Redis
MCP
Railway

Deployed · single-user · dormant since May 2026

204smoke-test assertions across five harnesses

Single-user by design — no public URL, no usage metrics.

Content updates ship with zero redeploys — a desktop session pushes concepts into the running bot.

REC-05Supporting

career-ops

A multi-stage application pipeline — discover, score, draft, adversarially verify, digest — with the submit decision kept human.

Multi-agent
Draft ⇄ verify
Deny-list

Personal-use system

103machine-checked integrity guardrails across five suites

Built for one user — me. Nothing auto-applies and nothing auto-publishes.

Red-teamed against a deny-list of claims I am not allowed to make about myself.

REC-06Supporting

This site

The site as its own case study: RED-first TDD with hook-enforced agent boundaries, an append-only decision log, and an accessibility gate with veto power over the design system.

Next.js
CSS Modules
RED-first TDD
axe WCAG 2.2 AA

Live

Two design decisionswere written by a failing accessibility check — a recoloured brand token and a scrapped opacity fade.

A portfolio site, not a product — the only thing it sells is the record.

The automations behind it are in section 03.

REC-07Supporting

Bootcampr

A platform that paired bootcamp graduates into cross-functional teams to ship real apps from PRDs over one-week sprints. Three and a half years, four iterations, the Antler accelerator.

React/TS
MERN
Webhooks
Google Calendar API

Closed October 2025

First industry rolesseveral team members landed them using the work as direct experience.

Closed publicly, with accountability, when the unit economics didn’t converge.

02The operating system

Two roles. Two files. One memory.

Every project on this page is built by the same system: a planning session that owns product thinking, specs and decisions, and a coding agent with exclusive write access to the code. They never share a workspace. They communicate through exactly two files.

FIG 2.1 — the two-role loop

Planning session

Product thinking · no write access to code

Scope & tradeoffswhat is worth building
Specsthe contract the code is held to
Decision logappend-only · one-line index

Two files · the only channel

Handoff brief →planning → code

← Return channelcode → planning

Coding agent

Exclusive write access · no product calls

Implementationthe only seat that edits code
Tests firstRED before green
Commit + gatered suite, no merge

No shared workspace. If it isn’t in one of the two files, it didn’t happen.

Material decision protocol

Any decision that changes a spec updates that spec in the same operation. The memory cannot drift from the decisions.

FIG 2.2 — a repo’s memory layer, split by load cost

Resident at boot · 2 files

Loaded before any work begins

Rulebook

How this repo works, and what a session is allowed to touch.

Manifest

A generated map of the whole corpus, described in a fraction of its size.

Fetch on demand · trigger-routed

Pulled only when named

Context

Current state; where the last session stopped.

Decisions

Append-only log under a one-line index.

Tracker

What is open and what is next.

Specs

The contract, updated in the same operation as its decision.

Handoff

Planning → code brief for the current unit of work.

Return channel

Code → planning report on what shipped.

Research dossiers

Sibling-project findings, refreshed on demand.

Scratch

Session notes, folded in at close.

FIG 2.3 — fetch router

Trigger vocabulary	Routes to	Target
decision, DEC-, locked, rationale	→	decisions/INDEX.md
handoff, coding task, ship	→	HANDOFF.md
rule, ritual, wrap up, cross-sync	→	MAINTENANCE.md
brand, palette, type, monogram	→	specs/brand-identity.md
research, case study, evidence	→	research/*

Routes resolve by trigger vocabularyRegenerated at wrap-up

FIG 2.4 — session ritual

BootLoad two files
The rulebook and the manifest. Nothing else.
AttestAttestation line
The session states what it loaded and what it may touch.
WorkFetch on trigger
Specs and decisions pulled only when named.
CloseWrap-up
Scratch folded in, decisions appended, manifest regenerated.

REC-03 · Flagship of section 02

CoBrain

Personal runtime · 3 repos onboarded

My cross-project context runtime, packaged as a Claude plugin — init, start, wrap, coord — and onboarded across the Portfolio, Ambito and OneLook repos.

Problem

Every session paid an always-on context floor: six-plus files loaded before any work began, most of them never used.

Mechanism

Boot loads a resident rulebook plus a generated manifest that maps the whole corpus in a fraction of its size. A fetch router keyed on trigger vocabulary pulls files on demand; wrap-up regenerates the manifest.

Heartbeats and advisory write-leases keep parallel sessions on one repo from colliding.

Result · portfolio onboarding−52%resident context

Nothing lost. Everything demoted to fetch-on-demand stays one trigger away.

Before: Six-file always-on context floor
After: Two-file boot plus a fetch-on-demand router

A personal system, not a product — no external users.

FIG 2.5 — cross-sync

Sibling-repo dossiers refresh on demand

Each project’s findings live as a dossier in the others’ memory layers, pulled forward when a session names them. That is what keeps this record current instead of annual.

The page you are reading is maintained by the system it describes.

FIG 2.6 — decision log · reversals

Reversal	Decision	Verdict
DEC-022 → DEC-031	TDD cadence reversed — the impl-first premise was refuted in practice, and RED-first restored by moving the blocking gate off pre-commit.	Reversed · locked
DEC-109 → DEC-115	Direct branch-merge-to-main reversed — a PR-based fleet workflow behind required CI.	Reversed · locked

Reversals sit next to what they replacedNothing is edited out

03Automations & skills

Most of these exist because something went wrong.

Grouped by the system each one serves. Where there is an incident behind a guardrail, the incident is on the record — the correction is the evidence, not the automation.

Portfolio site pipeline

Goal · no unverified code or claim ships

test-author subagent

Writes exactly one failing test, then hands back. Two PreToolUse hooks hard-block it from touching non-test files and from running git.

Origin on record

01Incident
It branched and committed on its own initiative, mid-build.
02Guardrail
The role was stripped to author-only, and the ban was enforced in hooks rather than a prompt: no non-test files, no git.
03Now prevents
A test author reaching implementation or version control. The hook carries its own regression test.

Gate-split TDD

Pre-commit stays fast and green while failing tests can still be committed first; the blocking test gate lives at pre-push and CI.

Origin on record · reversal

01Incident
An impl-first cadence was tried and its premise was refuted in practice.
02Guardrail
The reversal was locked as DEC-022 → DEC-031: commit RED first, block later.
03Now prevents
A pre-commit hook that punishes the RED step it is supposed to protect.

Accessibility gate

axe WCAG 2.2 AA runs on every PR across the core pages. This one has no incident behind it — it runs in reverse, and it has redesigned the design system twice.

Origin in reverse

01Verdict
It failed a brand colour’s contrast. Steel was recoloured — DEC-037.
02Verdict
It killed an opacity-fade entrance, replaced with a transform-only rise — DEC-038.
03Standing
The robot has taste veto. Two design decisions on the record were written by a failing check.

PR quality gate

Lint, types, tests and build all required behind a branch ruleset. A gitleaks secret scan runs alongside, an AI code review comments on every PR without blocking it, and Dependabot keeps the SHA-pinned actions fresh.

Blocking · review advisory

Session guardrail hooks

Write, bash, config and mutation guards, plus a format-on-stop pass — they police the coding agent’s session itself, not just its output.

PreToolUse · Stop

Slash commands

/go
/affected
/verify-app
/scaffold-route
/scaffold-component

Operator shortcuts

CoBrain runtime

Goal · memory that maintains itself

Manifest builder

Regenerates the corpus map, so the one file every session boots from is never stale.

On wrap-up

Coordination script

Heartbeats plus a generated recent-changes block, so two sessions on one repo can see each other.

Heartbeat · write-lease

Footprint check

Lints the resident-context budget. The two-file boot is enforced, not remembered.

Budget lint

Trigger evaluation

Tests the fetch router’s vocabulary against the corpus — a route that stops resolving is a failing test.

Router regression

career-ops pipeline

Goal · no application ships an overclaim

An automation whose job is to limit its owner’s claims.

A second agent attacks each draft, up to three iterations, and the whole pipeline is red-teamed against a deny-list of claims I am not allowed to make about myself. The interesting part isn’t the drafting — it’s that the verifier is adversarial by design: a draft that fails verification goes back for another iteration, not out the door.

Discover
Score
Draft
Adversarial verify
Digest

FIG 3.1103

checks a draft has to survive before it can leave the pipeline

Personal-use system

OneLook orchestrator

Goal · the build reaches a commit without a copy-paste step

Fire-and-forget agent dispatch

Agents are dispatched without a human waiting on them; an auto-commit chain and a file-based handoff state machine carry the work to a commit.

Origin on record

01Incident
An audit of the orchestrator’s own RCE exposure — it was auditing itself, and the finding was real.
02Guardrail
A cwd-jailed command tool with a deny-list, written from that audit.
03Now prevents
Commands reaching outside the repo, and handoffs that skip the commit gate.

FIG 3.2 — handoff state machine

Idle
Awaiting agent
In progress
Awaiting commit
Completed

FIG 3.3 — war story

The stdio-inheritance hang

The dispatched build agent hung at zero percent CPU, no error, nothing in the logs. The spawn had passed a working directory and nothing else — so the subprocess inherited the orchestrator’s stdin, which was a live MCP JSON-RPC protocol stream. The agent was sitting there reading protocol frames as its own input. The fix was three tokens of stdio config: stdin nulled, stdout and stderr piped for notifications and error detection.

DocumentedFixed

FIG 3.4 — war story

Sixty-second MCP timeouts

The MCP client enforces a sixty-second connection timeout; an agent build task takes two to five minutes. Every blocking dispatch died at the one-minute mark — the agent itself survived, but the caller had no way to confirm it had ever started. The fix inverted the contract: fire-and-forget dispatch that returns in about a second, with completion signalled through the handoff file and a macOS notification, and a watchdog timer that kills any agent that never exits.

DocumentedFixed

StudyAgent

Goal · content updates with zero redeploys

MCP content interface

A desktop AI session extracts concepts from course PDFs by vision and pushes them into the running bot. No redeploy, no migration.

Live content push

Job queue

Self-rescheduling — quiz pings land inside the study windows and re-queue themselves.

Self-rescheduling

Skills

Loaded before acting, not recalled mid-flight

Codified conventions the agents must load before they act — the difference between hoping an agent remembers the rules and making the rules part of its boot.

git

Branching, commit shape, PR hygiene.

Loads before any commit or PR

testing

Carries the TDD loop and the test-type matrix, so the cadence isn’t re-argued each session.

Loads before writing tests

nextjs-ui

Route, component and MDX conventions for this site.

Loads before touching routes

worktree

Parallel-branch setup, paired with the runtime’s write-leases.

Loads for parallel branches

Status chips are mandatory on every record. Ember marks live or deployed; steel marks paused, dormant, or personal-use. Numbers appear once per record and are sourced — a record with nothing to count carries a statement instead.

Also on the record

NextStepA cross-platform Expo/React Native daily-dashboard app: real Google OAuth, EAS builds, third-party API integration across two SDK iterations. Not shipped, no users, and no AI — it predates the AI-native work. Here as mobile breadth, not a flagship.
Old bootcamp projects (2021–22)Archive only. The old erickmanrique.com is my first portfolio; it’s stale and this rebuild replaces it. Not linked.