tcash.ca · field dispatches
The archive · every dispatch & reading
fredericton, nb filed_
tcash.ca · the archive

ARCHIVE.ca

Every dispatch and every reading, in one drawer.

The home page shows only what's newest. This is the full drawer: the dated dispatches in the ledger, and the library of readings and reference texts they're drawn from. Nothing here expires just because it stopped being new.

◂ Latest RSS / feed day · warm
§ 01 · The Ledger

Dated dispatches

11 filed
Showing 11 of 11 filed
31 MAY2026
Don't wait for Enter

A routine is just a Claude Code session that triggers itself. Field notes on trigger, context, and steerability, plus a work order builder you fill in and copy.

Field Notes8 min
29 MAY2026
The Verification Tax

The agent can produce more than you can responsibly inspect. The next skill is keeping verification cheaper than generation.

Reports12 min
19 MAY2026
Stop handing me the markdown your agent had to fight

Six months of defaulting to HTML for plans, reviews, and diagrams: what changed, what broke, and the one carve-out I keep. A reply to Thariq and Kurtis Redux.

Reports14 min
14 MAY2026
The CLAUDE.md I actually keep

Twenty-one rules went in. Five survived contact with real work, and one of them was “write less here.”

Config9 min
06 MAY2026
Four moves for a stuck agent

When the loop stalls, I reach for the same four corrections, in order, before I touch the prompt.

Workflow7 min
28 APR2026
Annotated PR review, in one file

An HTML diff with the reviewer’s notes pinned in the margin beats a comment thread every single time.

Code Review6 min
19 APR2026
Kimi K2.6, two weeks in

Where the cheaper agent holds, where it folds, and how its prompting wants to be shaped differently.

Field Test11 min
10 APR2026
Reading like Karpathy

Treat the agent as a study partner, not a vending machine. A reading method that survives long sessions.

Research8 min
30 MAR2026
How OpenAI actually uses Codex

Notes from how a shop that ships every day delegates to its agents, and where the humans stay.

Field Notes10 min
22 MAR2026
Meta-prompting, the Garry Tan way

Ask the model to write the prompt before you write the task. Why the extra hop earns its keep.

Technique5 min
14 MAR2026
The seven crossings

Seven points where a human still has to step into the agent’s loop, and what each one costs you.

Essay12 min
05 MAR2026
Learning on the shop floor

Skill compounds where the work happens, not in the docs. A case for letting the agent watch you work.

Essay9 min
§ 02 · The Library

Readings & references

16 on the shelf

The dispatches don't come from nowhere. These are the pieces behind them: field readings of the corpus I work from, and the reference texts they argue with. Undated by design; they're a shelf, not a feed.

Field readings 11 pieces
What the library already knows Twenty-four files on the shelf, read end to end and squeezed for what survives: the cross-cutting ten first, then by cluster, then a per-file ledger so nothing's dropped.
~24 min
Eighteen articles the corpus does not have yet A ranked accessions register of outside voices, counter-evidence, pricing data, and measured studies that patch the shelf's four self-diagnosed gaps.
~25 min
Where the techniques break The missing half of every how-to here: not how things fail without the technique, but how each technique itself fails. A boundary survey, written from inside the tribe.
~15 min
Inside karpathy/autoresearch A field manual for the overnight loop: how an agent runs ~100 experiments while you sleep, and the surfaces that keep it honest.
~20 min
Four surfaces, eight models Default surface vocabularies across eight models, and the behavioral connections between them.
~27 min
Twenty-one rules for the file Claude reads first CLAUDE.md, beyond the developer use case: the long version of the rules that made the cut.
~19 min
Four rules for an agent that codes The Karpathy-Skills specimen sheet: fat skills, fat code, thin harness.
~12 min
Getting the most out of Codex A close reading of how to brief Codex so it finishes the job instead of the first step.
~13 min
Using Codex goals effectively Goal mode, read carefully: PLAN.md as program, EXPERIMENTS.md as the ledger that survives.
~11 min
Driving the browser, four ways A comparative study of four ways to put an agent behind a browser, and where each one breaks.
~9 min
sqlc, explained Tools, demystified: what sqlc actually does to your queries, and why it earns its build step.
~6 min
Reference texts 5 pieces
The Unreasonable Effectiveness of HTML The source essay this blog argues with: Thariq's case for HTML as the default canvas between you and your agent.
~11 min
HTML effectiveness: specimen catalog The companion catalog: what the HTML-first outputs actually look like, specimen by specimen.
~6 min
Prompting best practices: Claude Reference: prompting guidance for Claude's latest models, kept on the shelf for lookups.
~35 min
Best practices for prompts: Moonshot AI Reference: Moonshot's own prompting guide for the Kimi models.
~10 min
Project instructions for “Format as html” The colophon piece: the project instructions that produce these very pages.
~19 min