How the system built itself
A timeline pulled from the workspace's own records — its git history, daily notes and decisions — not written from memory. The shape it makes is the argument.
The operator workspace has been logging its own construction the whole time. This is that record as one spine, from a build-to-learn habit years before any of the tooling through to a system that now composes itself. The interesting part isn't the list of things built — it's the shape: cold builds early, compositions later, which is what compounding looks like when you can see it.
Teal marks the composable machinery — engines and public surfaces. As the spine descends, entries start building on earlier ones: the shift from cold builds to composition.
The compounding gauge — computed from this data
50%
early eras
9 of 18 builds
78%
recent eras
29 of 37 builds
69%
overall
38 of 55 builds
The share of builds that compose earlier pieces rather than start from scratch — the honest measure of whether a system is compounding or merely getting bigger. It has risen from 50% in the early eras to 78% recently: the foundations were cold builds by necessity; almost everything since reuses them. These figures are recomputed from the timeline data on every update — if the number ever falls, something is being added without buying leverage.
Pre-history (build to learn)
Around twenty years working substantially with digital tools — a university coding course, then self-taught HTML/CSS in Dreamweaver and hand-built websites from ~2010.
Became a ChatGPT power user — custom GPTs, every model, multiple podcasts a week to accelerate learning.
builds onWeb-building roots
Ran substantial, multi-year projects entirely within ChatGPT — evidence-based, self-updating knowledge systems built by hand. Several were later systematised into code: the knowledge engine, the finance dashboard, the life-admin layer. The manual precursors to the operating system — the practice of building operating systems predates every tool.
builds onPower ChatGPT user
Started working in the terminal — GPT in one window, terminal in another — building basic games and simple calculators, pasting code from GPT's instructions.
builds onPower ChatGPT user
Adopted building-to-learn as the operating philosophy — the root of everything that compounds after it.
builds onProject-based ChatGPT work
Started using Claude Code when it entered the market (~Feb 2026) — the tool the whole operating system is built on.
builds onTerminal coding with GPT"Build to learn" philosophy
Built on Windows until ~March 2026, then moved to Mac — bringing only the worthwhile projects across via git. Git as a curated moving van; the first workspace commit is literally "migration prep."
builds on"Build to learn" philosophyClaude Code adopted
Era 1 — The practice takes hold (Mar–Apr 2026)
The governance practice predates the repo — the very first commit is already a "Session 3 close, migration prep." The practice, not a tool, is the true origin.
builds onWindows → Mac, via git
The first project scaffolded under governance — an AI commercial systems consultancy.
builds onOperator workspace on GitHub
A Next.js + YAML-backed household finance dashboard — an early standalone cold build (a parking-lot item shipped).
builds onProject-based ChatGPT workOperator workspace on GitHub
The review cadence begins — subtraction discipline and a monthly kill list.
builds onOperator workspace on GitHub
The memory layer begins — an Obsidian vault set up and the /capture skill rebuilt.
builds onOperator workspace on GitHub
The /diagnostic-lab concept and v1 spec; first brand identity.
An episode/bundle learning module plus /learn-extract (Firecrawl web data) — the first extraction engine.
An evidence-backed capability matrix (v0.1.4) and an audience-shaped profile module — the measurement layer.
The practice→tool crystallisation: 19 custom skills brought under version control in one move.
A weekly-review memory-promotion step feeding the Obsidian durable layer — concepts/patterns/decisions.
builds onObsidian vault + /capture
Pause/resume and auto-timestamping built into the session discipline.
CI wired across projects; parking-lot review; memory formalisation.
Era 2 — The RAG build (May 2026)
The operator-workspace RAG — corpus pivot, Phase 0, Layer 1 (embeddings + LanceDB over the workspace's own notes).
Eval set v0.1 + runner + first baseline — retrieval made measurable, not vibes.
builds onRAG corpus pivot + Phase 0/1
The /rag query skill; the smoking-gun question flips to YES; Phase 2.5 with a kill criterion locked.
builds onRAG corpus pivot + Phase 0/1
Push-as-the-public-gate added to the close routine.
The centrality decision: a usefulness check on every build + a website as the external vessel.
Phase 3 — a chat panel as the dashboard's 9th tab with LLM synthesis; the RAG becomes usable.
builds on/rag skill + Phase 2 done
Session-close auto-reindex (governor Step 12) — the corpus stays current for free, every close.
A voice / anti-jargon skill lifted from the chat-panel synthesis work.
The Universal Document Triage layer (/intake) — the filing clerk before everything — plus /skills-export audience modes.
Era 3 — Reusable engines (late May–Jun 2026)
A productisable per-corpus RAG module + read-only Gmail pull + /luna; the local-only storage mode designed in from day one.
builds onProject-based ChatGPT workRAG corpus pivot + Phase 0/1
A second corpus (learning-system) on the same module — reusability proven, not just claimed.
builds onLearning-system + /learn-extractKnowledge Inbox OS engine
Pilot v2 PASS — the RAG value case closed with real usage evidence.
Era 4 — The publishing flywheel (Jun 2026)
Scoped→built the same day; the portfolio/capability wing; a primary-slot succession.
builds onSkills matrix + profile moduleOption C — the external vesselKnowledge Inbox OS engine
Deployed to Vercel (auto-deploy on push) + the /promote skill (publish + sync) — the publishing flywheel engine.
builds oncommercial-builder website built
Six external-input skills hardened against injection; skills-matrix domain 11 → Proficient.
A paying-client external build brought in — and the external-project-intake convention it forced.
shared/video/ (engine/skin/content) + two live videos — a second brand forced motion into a reusable setting.
A second external-build run — a self-initiated alternative-folk artist concept (intaken 06-16; repo history predates it).
A brand-voice marketing-asset skill.
Set up and launched — the distribution/synthesis arm of the publishing flywheel.
builds onWebsite LIVE + /promote
Era 5 — Defensible Work Position phase (late Jun–Jul 2026)
A slot succession — the phase becomes the primary build; the site + Substack move to maintained.
A decision-support orchestrator and the advisory-engine module behind it.
BD meeting-prep with a verification layer that refuses to launder single-source claims — the phase's flagship.
builds onKnowledge Inbox OS engineDefensible Work Position activated
A registry + commit-owner rule so three concurrent windows don't clobber each other.
A cv-best-practices overlay and an evidenced, repositioned 2-page CV.
The first post-launch cadence post (trust-boundary).
builds onOperating AI-First Substack
Telegram → _inbox capture (text/photo/doc), always-on via LaunchAgent — capture from anywhere.
builds on/intake + /skills-export
A JD↔capabilities matcher (later gaining a cover-letter mode) with an honesty layer.
The decision to target AI-first companies, plus two CV variants (AI-first default, transformation fallback).
A read-only health command wired into session start/close (out of a Codex stability loop) — drift becomes visible.
Own domain live; the AI-first LinkedIn About live; /role-fit gains a cover-letter mode.
builds onWebsite LIVE + /promote
A Credentials RAG and a grounded, cited "Ask about Myles" CV chat LIVE at /cv — the CV becomes a system.
builds onKnowledge Inbox OS engineWebsite LIVE + /promoteDefensible Work Position activatedRepositioned 2-page CV
A Codex reviewer → inbox → Claude second-pass → user-approves loop; JD-fit-check on /cv; the British Land flagship demo.
builds on/meeting-brief (flagship)AI-Interactive CV + Credentials RAG
All five phases done; maintained going forward via the career-maintenance loop.
builds onDefensible Work Position activated
Era 6 — Market Intelligence Engine + fortify (Jul 2026)
A new engine composing deep-research + Knowledge Inbox OS + a thin new layer — the first fully-automated real weekly report. The current primary build.
builds onKnowledge Inbox OS engine/meeting-brief (flagship)
A 10-agent audit executed same-session: every-week fabrication gates, a privacy-of-descriptions scrub, a live WAF rate limit, encrypted backups, secret perms.
builds onworkspace-doctorWeekly Codex automation loop + JD-fit
Adzuna+Reed discovery → Claude /role-fit scoring → Telegram delivery, weekly-scheduled and proven autonomous — a passive market radar composed from existing parts.
builds onmobile-capture service/role-fit
A structured encrypted vault + master-PDF export + native doc-search over 223 documents (local-only).
builds onProject-based ChatGPT work
The full discipline loop narrows v1.2 to a trusted spine; plus an iCloud offsite backup leg, restore-proven.
builds onMarket Intelligence Engine (Phases 1–3)
Substack #3 + five site pieces + 39 OG cards + the mobile-capture watchdog — the whole site given a visual-identity layer.
builds onWebsite LIVE + /promoteOperating AI-First Substack
A posting decision, the weekly review, the June archive — and this build-timeline itself: a byproduct of governance already run, which is the compounding thesis proving itself.
This page — the curated record rendered live, with a compounding gauge computed from its own data on every update. Built the same day as its data, by composing the site's existing diagram registry and publishing pipeline: timeline-worthy by its own rules.
builds onWebsite LIVE + /promoteLinkedIn decision + weekly review + this timeline
A SQLite entity spine under the intel engine — append-only revisions, an auditable merge ledger with an exact-undo split path, an adversarial eval gate (zero false merges), byte-identical shadow parity, registry-driven ingestion — all four phases in one day; cutover still gated on the parallel-run verdict.
builds onMarket Intelligence Engine (Phases 1–3)MIE v1.2 scope BUILD-READY + backup leg
A reusable claim-verification layer — ranked materiality, adversarial escalation, a verdict ledger as the interface — unifying the three hand-built verify disciplines already running in sauna, market-intel and meeting-brief.
builds on/meeting-brief (flagship)
Compose → fan out across Claude/Perplexity/OpenAI → verify via /verify-claims → consolidate. Validated at two legs on real meeting briefs, where cross-model divergence forced primary-source verification.
builds on/meeting-brief (flagship)/verify-claims
The engine completed at three legs and ran end-to-end for the first time: 24 claims live-verified, three killed (including a £10,000 price error), and the resulting consumer guide published live — every pipeline layer caught errors the previous layer produced.
builds ondeepreach multi-model engine
Three read-only-first operator tools built in one morning: a where-am-I snapshot composing the doctor and governance files, a machine health report, and a propose-then-approve file-hygiene pass — the tidy scanner improved twice through first contact with the real workspace.
builds onworkspace-doctor
A real work job exposed the gap: three verification tools, three different vocabularies for how solid a fact is, and no written source hierarchy. One evidence scale, one ranked source hierarchy (registry over firm page over snippet; auth-gated proves nothing), a deployment map, and /roster-verify for the shape none of them covered — checking a whole list of entities rather than one claim or one company.
builds on/meeting-brief (flagship)/verify-claimsdeepreach multi-model engine
Era 7 — The commercial front door + the reliability turn (Aug 2026)
An agent rebuilt on the Claude Agent SDK as an autonomous second opinion on the diagnostic chain, then compared in writing against the file/prompt approach it had been preferred over for months. The conclusion was only reachable by building it: the two compose rather than compete — the SDK runs the skills already written — and the real dividing line is enforcement. A skill can say "don't write files"; only a harness can prevent it.
builds onDiagnostic Lab spec + brand
The market-intelligence engine's four-week silent parallel run reached the date it was set against and returned a split answer: production cuts over, circulation does not — blocked behind six named defects. The verdict was then amended three times against its own author as the evidence was re-read. A gate designed in advance, run to its date, and allowed to say no.
builds onMarket Intelligence Engine (Phases 1–3)MIE v1.2 trusted spine (all 4 phases)
A standing guardrail against any selling surface was relaxed deliberately rather than drifted past — services described, no price published, buyer-first, the CV demoted but kept. Recorded as a falsifiable judgement with a dated test attached: one genuine enquiry and one offer change by 30 November, or the front door is decoration.
builds onAI-first positioning + CV variants
A complete consent and disclosure layer shipped on the public site — analytics gated behind an explicit choice, with privacy, cookie and terms pages published to state what it actually runs under. The absent dismiss button is the decision: no dark-pattern default-yes.
builds oncommercial-builder website built
An idea filed with no way back is not filed, it is buried, and it looks perfectly filed the whole time. Six checks over the idea library now run at every session open and close, surfaced on demand, with the capture step as a preventive gate that will not file an idea without a route back. Two real defects on day one, including a note that had been pullable and unnoticed for ten days.
builds on/intake + /skills-exportworkspace-doctor
Opening a session used to propose one thing — the carried re-entry point. It now returns every strand that could legitimately be worked, numbered, each with where it stands and its next concrete step, plus a mandatory one-line recommendation. Built because the capability existed elsewhere and sat outside the reading path at the exact moment it was needed.
builds onParallel-session governanceNothing captured gets lost — the route-back guarantee
Work driven from a phone after leaving the desk, which only holds if the machine stays awake — this one sleeps after a minute idle. A keep-awake tool built to be asked in plain language, honest about the one thing it cannot do (closing the lid still sleeps it) and about verifying its own claim against the operating system rather than its own bookkeeping.
builds onHealth & hygiene tool cluster (/status · /mac-health · /tidy)
Owner decision on the market-intelligence engine: paid sources permanently out of scope, and — because the old circulation gate required beating a paid source, which becomes unpassable once paid sources are dropped — the gate was replaced rather than waived. A week's output judged useful on review, on open sources, with the coverage gap disclosed in the report itself. No further build work scheduled.
builds onMIE v1.2 trusted spine (all 4 phases)
Entry paths rebuilt around what a buyer is trying to do rather than what exists, the systems regrouped onto a four-system frame, and a services page shipped that routes into existing proof and publishes no price. Three plan stages in one evening.
builds oncommercial-builder website builtmylesmellor.co.uk + LinkedIn AboutThe commercial front door opened by decision
A public form whose submissions are classified against an ideal-client profile by a model at the moment they arrive, stored already-triaged, and notified with the verdict — cost sized before building and capped, failing safe so a triage error stores the enquiry untriaged rather than dropping it. The first surface here that runs unattended on a stranger's input rather than being a tool its author operates.
builds onThe commercial front door opened by decision
A one-hour Claude Code course delivered 1:1 to a colleague, with a starter pack left behind so it can be re-run alone, plus the session's finale ported into a reusable component kit. Notable for what it is rather than its size: the first item on this timeline where something built here was handed to another person and used by them.
builds onLearning-system + /learn-extractAgent SDK comparison — the choice tested instead of asserted
Research kept hitting sources that were PDFs and quietly scoring them as unavailable. A small fetch-and-extract tool closed the gap — confined output, magic-byte checking, sanitised filenames, private-network refusal — and then earned a second pass when a scanned PDF returned silence rather than saying it was scanned.
builds on/verify-claimsVerified-research standards + roster verification
The multi-model research engine gained a third independent provider, probed before building rather than after, and built to fail loudly instead of silently falling back to an ungrounded answer — because a research leg that quietly stops being grounded is worse than one that is absent.
builds ondeepreach multi-model engine
A production pipeline was proven non-deterministic by a purpose-built probe rather than inferred to be, and every model call in the workspace was swept — then the fix failed to generalise, because the model behind one feature has no such control at all. The output was a reporting standard rather than a code change: any eval scored once against a model with no determinism control carries an unstated error bar, including this workspace's own previously-published numbers.
builds onRAG eval harness
I didn't set out to build a timeline. The workspace produces one as a by-product — every session closes with a commit, every decision leaves a record, every capability lands with a date attached. Curating those into a single spine took an afternoon, because the data was already there. That it exists at all is the first piece of evidence for what it claims: a system that documents itself as it runs is one that compounds rather than just accumulates.
It didn't start with a tool
The instinct is to say it began with the first script. The record disagrees. The very first commit is already a session close — the governance habit (working in disciplined sessions, closing them properly, keeping the files honest) predated the repository by weeks. Further back still, the real origin isn't a tool at all: it's a decision to build things in order to learn them, and years of running long, evidence-based projects by hand before any of this was code.
That's why the earliest part of the timeline is drawn dashed. Those entries are recalled, not recorded — and a timeline whose whole job is to be true shouldn't render an approximate memory with the same confidence as a dated commit.
The shape is the argument
Read the spine top to bottom and the same signal keeps appearing: early entries stand alone, later ones build on the entries above them. The foundational engines — the retrieval layer, the knowledge module, the video pipeline — had to be cold builds, because nothing existed yet to compose from. Everything since increasingly reuses them. The market-intelligence engine, the interactive CV, the job scanner: none were built from scratch; each is a recombination of pieces already on the board.
There's a number under this. Of the capabilities on the timeline, a little over half compose earlier ones rather than starting fresh — and that fraction is rising, from about half in the early months to roughly three-quarters lately. That ratio is the honest measure of whether a system is compounding or merely getting bigger. If it ever starts falling, something is being added without buying leverage. Right now it climbs.
The rules that keep it honest
A curated timeline is a place where a story can quietly become tidier than the truth — drop the dead ends and every move looks like it landed. Three rules guard against that. Every node links to a source, so nothing flattering can be smuggled in that didn't happen. The pre-history is marked approximate. And the projects that seeded this — the ones I'd been running by hand long before the workspace — appear only as their shape: a single node showing that hand-built work became coded systems, never a word of what those projects actually contained. Some of them hold private things. The timeline shows the pattern and keeps the contents; that distinction is the same one the whole system is built on.
The artifact that demonstrates the compounding turns out to be able to measure it too. That was the point.