How the work gets done

Sessions

See complete work sessions, including mistakes, corrections, and the evidence behind each decision.

Open a full session or jump to one tested technique you can reuse.

18Sessions 11Techniques

Session 02

Start with a complete sessionThe Browser Is a Shell CommandOpen the session

01See the work as it happened.Follow the decisions, failures, and corrections in order.

02Reuse a tested technique.Take one useful part without replaying the whole session.

Complete collection

Browse sessions and techniques.

Search all 29 sessions and techniques, or open either collection directly.

29 of 29 sessions and techniques in view

All sessions and techniques

Full sessions

18 sessions

See the work in the order it happened.

Published Instagram story rendered by the demo's pipeline

delegation · 2026-07Twelve Messages

Everything typed to take a live event from spreadsheet chaos to published social content — and the method that made the other 99% happen.

The verbatim conversation, a five-principle method, the pipeline that runs it, and two honest failures.

Open session

Demo 02 title slide: the browser is a shell command, with a real browse-shot transcript

tools · 2026-07The Browser Is a Shell Command

Replacing 18,000 tokens of always-loaded browser-automation schema with ten shell commands and a README — and why the agent gets sharper, not weaker.

The cost math, four design choices that make a tool agent-friendly, live transcripts, and where MCP still wins.

Open session

Demo 03 title slide, featuring a real read-guard denial of an oversized screenshot

enforcement · 2026-07Taught Once, Enforced Forever

Corrections given in chat decay when the session ends. The fix is changing the agent's environment — helpers, deny-hooks, and CI ratchets that carry the rule forever.

Four real guards with their origin audits, verbatim deny messages, and the five properties that keep guardrails from getting ripped out.

Open session

Demo 04 title slide: a real artifact-miner run over 2,747 transcripts

memory · 2026-08Your Sessions Are a Corpus

A 24 GB session archive forced the next question: how do you keep the lessons, prove closeout, and let the raw record expire?

The mining loop end to end, plus the measured retention incident that separated canonical truth, compact recall, raw transcripts, and reproducible worktree output.

Open session

Demo 05 title slide with a real agent-lane ticket and its labels

autonomy · 2026-07The Product That Files Its Own Tickets

End users file feedback, an LLM judge triages it into GitHub Issues, an agent implements the safe ones — with the autonomy boundary in deterministic code and a human holding the only merge key.

The four-stage loop, the allowlist gate read from live source, six real tickets sorted by lane, and a run ledger whose best result was declining to fabricate a fix.

Open session

Demo 06 title slide with the shipped-twice landmine exhibit

knowledge · 2026-07The Registry of Landmines

Some load-bearing facts are invisible to search — the code compiles and the behavior is still wrong. One registry file holds them, a derive catalog enforces them in CI, and a meta-test keeps the registry from lying about itself.

The bug that shipped twice, the token economics of pre-paid conclusions, the why/decided/currently-true split, and the registry-about-the-registry.

Open session

Demo 07 title slide with a claim-versus-runtime-check exhibit

verification · 2026-07The Agent Said It Checked

"Verified" is a sentence, not a fact. A security migration passed its audit while production was broken — and the discipline that came out of it caught three more false claims building this very series.

The circular-audit failure mode, a prod-broke-twice case study peeled layer by layer, and the re-derivation discipline with its real budget.

Open session

Demo 08 title slide: the staged pipeline with gate posts between every stage

process · 2026-07Gates Between Agentic Stages

Agents are strong inside a stage and unreliable at the boundaries. A delivery methodology built on that: deterministic gates between agentic stages — 88 versioned revisions, 14 running initiatives, and the day production published fiction.

The deterministic-core/agentic-shell shape, the migration-sweep incident and the inverted check it left behind, and a methodology versioned like a product.

Open session

Demo 09 title slide: a README claim beside the code fact that contradicts it

documentation · 2026-07The Beautifier Was an Auditor

A skill built to make a README prettier turned out to be a drift detector. Across five repos it flagged a deploy doc still pointing at a platform the project left months ago, a command list that no longer matched the actual commands, and capabilities the code had grown that nobody had written down. The visual polish was the smallest thing it did.

Why a beautifier's real payload is its inspect step, five repos' worth of drift it surfaced, and one honest failure — claiming checked when the check was only partial.

Open session

Demo 10 title slide with a confidence tag being downgraded from verified to search-derived

applied advisory · 2026-07The Chiropractor's Four Questions

A sports chiropractor asked four plain questions about selling rehab programs from his website. The AI's answer was confident, tabulated, sourced — and its evidence was marketing all the way down. Twice.

A first pass built on vendors ranking themselves, a 'correction' that swapped in a competitor's stat, three refuted claims with the fetches that killed them, and the confidence-label system that came out the other side.

Open session

Demo 11 title slide: the sycophancy was in the config, with the operator's verbatim question about model versus configuration

config experiment · 2026-07The Sycophancy Was in the Config

A rival model reviewing my agent's work looked sharper — it demanded competing candidates and restraint. So my agent A/B tested its own harness: same model, same prompt, config on and off. The vanilla arm pushed back on me twice. The configured arm never did.

A twelve-run experiment on a live harness: the bare-mode control arm, a self-contaminating probe, a scorecard separating model habits from config damage, and one validated sentence that flipped the behavior 2/2.

Open session

Demo 12 title slide: one component I didn't already have, with the operator's verbatim request to compare the tool against what he already runs

tool adoption · 2026-07One Component I Didn't Already Have

A plugin advertising eleven agents and fifty-four hooks, measured against the setup I already run. Of the eight components that could land on my machine, three were already going, one was the wrong platform, one was broken upstream, two were costs. The delta was one.

A component-by-component subtraction against a real machine, three mechanical queries against the part of a repo nobody curates, a retracted recommendation, and the correction that produced all of it.

Open session

Demo 13 title slide: the gate was written first, with the preregistration commit that predates the model download

admission gates · 2026-08The Gate Was Written First

Three ways to make a meeting archive answer a question, three preregistered gates, three refusals in one day. Each gate was committed to git before the model weights finished downloading — so when the scores came back plausible, there was nothing left to negotiate with. The feature that shipped is the one the evidence admitted.

A one-day trilogy of preregistered probes — a similarity floor, a pairwise scorer, an extractive no-answer head — each gate frozen before first inference, each failed without tuning; the falsifier showing a right answer scoring below a wrong one, which no threshold can fix; and the honest feature that shipped instead.

Open session

Demo 14 title slide: Kibble & Co. doesn't sell anything, the fictional pet-supply storefront built to make subscription flows real

derived not narrated · 2026-08Kibble & Co. Doesn't Sell Anything

A pet-supply storefront went live on a real BigCommerce sandbox, moving real Stripe-sandbox money through real dunning logic — and it exists only so a subscription-billing platform could be built against a genuine store instead of a slide deck. Five deployed apps, 94 ADRs, and a generated file that answers 'what's shipped' so nobody has to narrate it. The method is the point; the pet food is the test fixture.

A five-app BigCommerce subscriptions platform — storefront, admin, API, docs, marketing — built against a live sandbox store behind a fictional pet-supply merchant; 94 ADRs and 218 user stories moving through one repo; the mechanical state-derivation file that replaces status narration; and two real failures, a circular verification audit and a GitHub rate-limit exhaustion, with the fixes that shipped.

Open session

Demo 15 title slide: the harder deliverable wasn't the build, showing the ten-document bid package's document list and story counts

vendor scoping · 2026-08The Harder Deliverable Wasn't the Build

A ten-document vendor bid package — the scoping and estimation kit a global systems integrator needs before it can bid on a BigCommerce-native subscriptions build — projected from a 28-epic, 218-story spec corpus in days, not weeks. A leakage lint hard-fails before a forbidden pattern ever reaches the export, and a bundling script freezes identical, name-stamped copies per vendor. Every document still carries status: draft-internal-review — the release checklist stayed a human's job.

How a 10-document, 28-epic, 218-story vendor bid package gets projected — not authored — from a canonical spec corpus, sanitized by a hard-fail leakage lint instead of a proofread, frozen into identical per-vendor bundles, and held behind an unchecked human release checklist before anything ships.

Open session

Demo 16 title slide: the agent that asks twice, showing the preview-then-execute turn card

confirm then act · 2026-08The Agent That Asks Twice

Ask BC is a production agent with 29 tools against a live commerce store — 22 that read, 7 that can change it — deployed split across a Next.js app and a Cloudflare Worker with a Durable Object per store. Every mutation needs two turns: a preview, then an explicit confirm. Months later, a sibling platform copied the runtime wholesale for its own merchant copilot — same Durable Object, same codemode sandbox, same two-turn writes — and pointed it at a different API entirely.

The two-turn confirmation pattern read straight from the tool code — nothing mutates on the first call, only on an explicit second call with the confirm flag set to true; the generative-UI block protocol that turns tool output into an inline React component instead of a wall of markdown; the Haiku-default, Sonnet-on-retry model split; a real security gap caught and fixed a month before the app's custom domain went live; and the decision record where a second product adopted the identical Durable-Object-plus-codemode runtime for its own copilot, pointed at an entirely different data plane.

Open session

Demo 17 title slide: the layout was inferred, not designed, with the four-persona quad — gatherer, hunter, researcher, gifter

inferred not designed · 2026-08The Layout Was Inferred, Not Designed

Aisles is a headless storefront that decides its own layout per shopper — 31 weighted rules feed a Bayesian inference engine that picks one of four personas, then an AI model generates the category page from a four-component vocabulary a Zod schema won't let it leave. Same URL, same catalog, four schema-valid layouts. On 2026-08-12 it was pointed at a real BigCommerce store for the first time — a brand-config entry, two environment variables, a harvested palette with file-line citations, and one honest bug: a client-side lookup that silently rendered furniture navigation over a pet-supply catalog.

The full pipeline read from source — 31 rules, a gatherer-biased Bayesian prior, and a softmax over four named personas (gatherer, hunter, researcher, gifter); the Zod schema that makes an AI-written layout provably valid instead of usually reasonable; the per-persona prompt language that actually reaches the model; the same-day integration against a live BigCommerce sandbox that took one config file because the catalog seam was already generic; a palette harvested from the storefront's own CSS with file and line citations; and the client-side brand-resolution bug the integration surfaced, with its fix.

Open session

Demo 18 title slide: the guardrail removed the storefront, with the four-step chain from direction to handoff to hardening to removal

steering across agents · 2026-08The Guardrail Removed the Storefront

I asked one agent to make an AI storefront look as good as the real one, then handed the work to a second agent in four words: “pick up where Claude ended.” Nineteen commits later the AI was gone — and every commit message was a safety word. Nobody typed “remove the model.” The tests stayed green because they were rewritten to expect it. What caught it was remembering what the live site used to do.

The verbatim direction that started it and the four-word handoff that dropped it; 27 agent sessions in one repository in a day, zero containing a typed instruction — every operator turn an auto-generated approval; the nineteen-commit verb ladder, where no single commit is wrong and the sequence deletes the feature; the green status line reporting “zero model requests” as an achievement for a product built on model requests; the contract test rewritten from listing model-authorized zones to asserting there are none; and the second overshoot, where the agent sent to fix the mistake made it again first.

Open session

Techniques

11 techniques

Tested parts of the work you can reuse.

Applied techniqueTwo Ways to Draw a System

Hand-authored SVG vs themed Mermaid — chosen per diagram, both shown working.

Open technique Applied techniqueMine Your Own Transcripts

Treat your agent sessions as a searchable corpus — corrections become priors the next session reads.

Open technique Applied techniqueGate Every Hand-Off

Agents work inside a stage; deterministic gates verify between. Gate the seam, not the middle.

Open technique Applied techniquePromote a Correction to a Guardrail

Wrapper → deny-hook → CI ratchet: promote a repeated correction to the lowest rung that holds.

Open technique Applied techniqueA Registry for What Grep Can't Find

Facts grep can't surface, held in one registry and enforced in CI — the conclusion, pre-paid once.

Open technique Applied techniquePut Provenance Labels on AI Research

Three label tiers, an adversarial re-derivation loop, and the honest downgrade — shown working.

Open technique Applied techniqueBare-Arm Test Your Agent Config

A control arm for your own harness: bare mode, isolation arms, and rules validated before adoption.

Open technique Applied techniqueRun the Subtraction Before You Install

A feature list is half a subtraction. Enumerate your baseline, sort the tool's list into three buckets, then search its tracker for the uninstall.

Open technique Applied techniqueClose the Session, Keep the Lesson

Turn a finished agent session into durable project truth and compact recall before its raw transcript expires.

Open technique Applied techniqueOne Heavy Job at a Time

Local inference makes scheduling an operational decision — and a panic's exit reason is two published integers, not a guess.

Open technique Applied techniqueThe Home AI Server Question

A measured local-versus-frontier comparison, plus the cost inputs for deciding whether to buy a home AI server.

Open technique

Loading the complete collection…