Field report · Monday 22 June 2026

The thinking it shows you isn't the thinking that shipped

Three releases this week circle one problem: a build agent will give you its account of the work, and that account is not the work.

A Pop Art screenprint panel: a thick-outlined magnifying glass held over a fanned stack of file cards marked with small red and blue diff bars, a yellow comic starburst behind it, and a field of red Ben-Day halftone dots, all in flat primaries with heavy black keylines and no readable text.
Plate I · Inspecting the editsBen-Day, keyline, no chrome

Agentic tools spent the last year learning to act on your project. This week the work is about accountability: a tool that logs what the agent changed, a close reading of what its visible thinking actually is, and a survey of how teams are configuring the agents in the first place. Read together, they describe a week spent learning not to take the agent at its word.

Today's Art Direction

Pop Art Screenprint

The briefing as a comic-panel product page: flat primaries, thick black keylines, Ben-Day halftone fields, and speech bubbles doing real label work.

Pop art took commercial print, the comic panel and the screenprinted poster, and made the mass-produced surface the subject. Translated to the web it becomes a product feature page: a hero showcase with the plate framed in a heavy keyline, the news set as numbered feature blocks the way a landing page lists features, and labels carried by speech bubbles and stamps rather than quiet gray chrome. Ben-Day dot fields supply texture and the primaries do the zoning, while the reading column stays a calm flat field at 19px so the comic loudness never touches the words.

Ben-Day dotsKeylineSpeech bubbleFlat primariesHalftone fieldFeature blockStarburstScreenprint
Tooling · Provenance

A local trail that remembers every file the agent touched

Ponytrail is a small open-source tool that sits alongside an AI coding agent and records each change it makes to your files as a local history tree inside a .pony-trail folder in the project. For every edit it keeps the action, a summary, the result, and a way to revert a file to an earlier snapshot, all on your own machine rather than buried in the agent's session log.

It is a modest utility with a pointed premise: the agent's edits are something to be recorded and rolled back, not trusted on sight. The trail stays local, so the record of what happened does not depend on the tool that did it.

github.com
Technique · What you're reading

The thinking your coding tool prints is a summary, not the trace

Patrick McCanna went to inspect the reasoning Claude Code writes to disk and found an encrypted signature where the thinking blocks should be. The documented behavior fills in the rest: the model's full reasoning is encrypted, the key stays with Anthropic, and what your machine receives is a summary of the thinking rather than the reasoning that drove the session. Getting the real trace, he notes, requires an enterprise agreement.

That is the asterisk on last week's good news. The tools that started showing their work are worth having, but a visible thinking pane is a retelling, not evidence, and it should not be mistaken for an audit trail.

patrickmccanna.net
Workflow · Field data

Almost everyone ships a config file, far fewer split the work

A scan of public Claude Code repositories found that roughly 85 percent now ship a CLAUDE.md, the plain rules file that steers the agent, while only about a quarter use subagents to divide a job across specialists. The same survey watched the config files bloating: a tenth run past 24 kilobytes, the top one percent past 62, and the largest single file weighed in at 341.

The instruction file has become table stakes, in other words, but the structural moves around it are still rare, and a heavier rules file is not a better one. The lever most teams are not pulling is dividing the work, not writing more rules for one agent to ignore.

buildthisnow.com

You can't read the agent's mind. Read what it changed.

Prompt Lab

Recreate today's art direction
Pop Art Screenprint Product Page
Design a single web page as a Pop Art screenprint product feature page presenting a short briefing of linked stories. Page archetype: a comic-panel landing page that showcases one hero plate and then lists the news as product features, not a dashboard, blog, or card grid.

Ground: a bright warm screenprint cream (#F5ECCF) with a lifted panel (#FBF5E2). Text is ink (#1A1714) at 19px, line-height 1.68, measure near 64 characters, never justified. Use flat pop primaries: pop red (#E0301F) for marks and the Tooling zone, cobalt (#1E5AC8) for the Technique zone, pop yellow (#FFC20E) for the Workflow zone, hot pink (#E11C7C) used once as accent. Inline links use a deeper red (#B5180B) so they stay WCAG AA on cream; never set body text in yellow or bright red. Keep all color in marks, fills, and dot fields, not under the words.

Layout: a product showcase. A hero splits into a Bungee headline, deck, and a spec strip of metadata cells with thick keylines on the left, and a framed hero plate on the right backed by a red Ben-Day halftone field. The news runs as three numbered feature blocks (01, 02, 03), each opened by a zone-colored top bar, a halftone number bubble, a solid-filled speech-bubble kicker, a linked Bungee headline, one or two plain paragraphs, and a dashed source stamp. One giant display pull-quote sits on a yellow halftone field. A product-info card carries the art-direction note.

Type: Bungee for the hero headline, feature headlines, numbers, and the one pull-quote; Mulish for body and the single italic art-direction summary; DM Mono for labels, the spec strip, stamps, and the prompt block.

Components: speech bubbles and stamps as real labels (not dead chrome), Ben-Day dot fields for texture, thick 2.5 to 3.5px black keylines, flat color zoning by section. Include one generated plate: a comic magnifying glass over a fanned stack of file cards marked with abstract red and blue diff bars, a yellow starburst behind it, and a halftone field, all flat fills with heavy outlines and no readable text inside the image. Add one restrained micro-interaction, a link underline that flips from yellow to red on hover, behind a prefers-reduced-motion guard. No gradients, no glow, no neon, no pill-shaping everything, no accent left-border stripes, no fake search fields or toggles, no emoji.

Works in: Beaver Builder AI, v0, Lovable, Cursor, Figma Make.

Field Note

One thread runs through the week: a build agent narrates the work, and the narration is not the record. A tool that logs the real edits, a look at what the visible thinking really is, and the Verge's tally of vibe-coded apps that shipped with security holes nobody noticed all point the same way, toward checking before you trust.

Sources

  1. Ponytrail, a local audit trail for AI coding-agent editsgithub.com
  2. The text in Claude Code's Extended Thinking output is not authenticpatrickmccanna.net
  3. The State of Claude Code 2026buildthisnow.com
  4. Read this before you vibe-code another apptheverge.com