<- all tokdocs
2026.10.07@breakingagents

Dating every decision in a design doc is ADR practice, and adr-tools has auto-stamped a `Date:` field since 2016

Coding
Watch on TikTok

View on TikTok ->

The clip runs 109.269 seconds at 720x1280 (9:16, 30 fps) in an MP4/QuickTime container, carrying HEVC video (hvc1 tag, Main profile, level 120, yuv420p, 241,597 bps, 3,276 frames) and HE-AACv2 audio (44.1 kHz stereo, 48,083 bps, 2,357 frames), for an overall 296,225 bps and a 4,046,027-byte file, pulled as yt-dlp format bytevc1_720p_296222-1. All container and codec values come from ffprobe on the downloaded video.mp4. TikTok's metadata puts the post at 2026-10-07 16:15:18 UTC from uploader breakingagents, channel nickname "Breaking Agents", with audio credited to "original sound - Breaking Agents" and a 324-character description. At capture on 2026-10-08 the counters read 588 views, 25 likes, 1 comment, 3 reposts, and 11 saves, roughly thirteen hours after posting. I read the full 255-word transcript and 23 of the 55 extracted frames (frames 1, 4, 8, 12, 16, 19, 22, 25, 28, 31, 34, 36, 37, 40, 42, 43, 44, 45, 46, 50, 52, 53, 55, sampled evenly across the 2-second grid). The visuals alternate between a talking-head shot and a graph-paper title card. The speaker sits at a desk in front of two monitors, holding an orange RØDE foam-capped handheld mic, wearing a plain black tee. The left monitor shows a dark marketing page reading "Understand how to run any repository." with an orange accent button; I could not identify that product from a web search and leave it unnamed. Burned-in captions track the audio one phrase at a time ("let's talk about", "UI decisions", "that's the one", "change it"). The card overlays are monospaced and build line by line: "MY DESIGN DOC HAS DATES." over a blue-boxed decided 2026-09-16; "A MAC MENU-BAR APP" over a drawn menu bar reading 10:24 with an orange dot appearing on the active icon and the label "something to do"; "NEXT SESSION · EMPTY" filling in sees orange dot, proposes "make it red", decided × 2 days ago; "EVERY DECISION · 3 PARTS" filling in decision red only, reason disk 90% full, date 2026-09-16; a file listing headed docs/DESIGN.md 13 decided with thirteen dated one-liners from 09-12 Name 3 risk levels to 09-23 Design for light mode, with 09-16 Red only at 90% disk boxed in blue; "WHY THE DATE WORKS / 1985"; a "NAUR, 1985" card building to "A PROGRAM IS THE THEORY IN ITS BUILDERS' HEADS. / THE CODE ALONE CAN'T CARRY IT."; "WITH AGENTS, A BUILDER LEAVES EVERY SESSION. / THE DATE STAYS."; and a closing "YOUR RULES FILE" card reading one rule × no date × no reason → that's the one. An orange "+ Follow" pill appears over the talking head near frame 46.

The auto-transcript mangles the one rule the whole video is built around

Whisper renders the example rule as "Read only when the disk is full. Decided September 16." The on-screen text settles it three different ways. The "EVERY DECISION · 3 PARTS" card reads decision red only and reason disk 90% full. The docs/DESIGN.md listing reads 09-16 Red only at 90% disk. The rule is about reserving the color red for a 90-percent-full disk, not about read-only mode. Anyone quoting this clip from a transcript alone will get the example backwards.

Nygard's original 2011 ADR template has no date field, and the video's format is closer to adr-tools than to the canonical template

The video presents "decision, reason, date" as a personal discovery. The practice has a documented lineage with a specific order. Michael Nygard published Documenting Architecture Decisions on 15 November 2011, opening with "One of the hardest things to track during the life of a project is the motivation behind certain decisions." His template is Title, Status, Context, Decision, Consequences. It has no date field. The date arrived through tooling. Nat Pryce's adr-tools ships a template.md whose second line is literally Date: DATE, and src/adr-new fills it with date=${ADR_DATE:-$(date +%Y-%m-%d)}. That repo's own ADR 0001 is dated 2016-02-12, and ADR 0008, "Use ISO 8601 Format for Dates", dated 2017-02-21, states the reasoning directly: "An important component of the history is the time at which a decision was made." The video's 2026-09-16 and 09-16 are that same ISO 8601 choice.

The community standard caught up later. MADR reached 4.0.0 on 17 September 2024 (release date read from the GitHub API on adr/madr), and its template front matter lists date: {YYYY-MM-DD when the decision was last updated} as an optional field. Note the divergence. MADR's date records when the decision was last updated. The video's date records when the decision was originally made. Those answer different questions, and the video's version is the one that stops re-litigation. Joel Parker Henderson's widely used ADR collection states the general rule in its README: "Timestamps: Identify when each item in the ADR is written. This is especially important for aspects that may change over time, such as costs, schedules, scaling, and the like."

Naur's 1985 paper is titled "Programming as Theory Building," and his own conclusion argues against the video's remedy

The video says "Peter Naur. Programming is theory building," and the card reads "A PROGRAM IS THE THEORY IN ITS BUILDERS' HEADS." Reading the paper itself (via the reprint hosted at gwern.net, introduced there as written in 1985 and reprinted in Naur's collection Computing: A Human Activity, 1992), the title is "Programming as Theory Building," and the builders'-heads line is a paraphrase rather than a quotation. Naur's actual sentence is that "an essential part of any program, the theory of it, is something that could not conceivably be expressed, but is inextricably bound to human beings." He borrows the word "theory" from Gilbert Ryle's The Concept of Mind, defining what a programmer possesses as "a theory of how certain affairs of the world will be handled by, or supported by, a computer program."

The video's "when the builders leave, the theory dies" compresses a more specific claim. Naur writes: "The death of a program happens when the programmer team possessing its theory is dissolved." Death is dissolution of the team, not the departure of one person, and Naur adds that "a dead program may continue to be used for execution in a computer and to produce useful results."

The sharper problem is that Naur's conclusion points the other way from the video's fix. He writes that "program revival, that is reestablishing the theory of a program merely from the documentation, is strictly impossible," and that a new programmer needs "the opportunity to work in close contact with the programmers who already possess the theory." The paper argues that a document cannot carry the theory. The video cites the paper as support for a document that carries the theory. The dated line in docs/DESIGN.md is better described as a cheap guard against one specific failure, which is an agent re-opening a settled question, than as a solution to the problem Naur named.

The empty-context premise checks out against Anthropic's own docs, and explains why the file alone is not enough

The video's premise is that each agent session starts with no memory of the last one. Anthropic's Claude Code memory documentation states it in the first line: "Each Claude Code session begins with a fresh context window. Two mechanisms carry knowledge across sessions: CLAUDE.md files ... [and] Auto memory." The same page adds the caveat the video does not mention: "Claude treats them as context, not enforced configuration. To block an action regardless of what Claude decides, use a PreToolUse hook instead." A dated rule raises the odds the agent drops the argument. It does not guarantee it. For decisions where re-opening the question is actually expensive, the first-party guidance points at a hook rather than a line in a markdown file.

Several of the video's supporting claims cannot be checked

"I'm a staff software engineer," "I have built two applications with coding agents this year," and "20 minutes of going back and forth in the prompt" are first-person claims with no public artifact behind them. The docs/DESIGN.md screen is a stylized graphic rather than a screen recording of a real file, so the "13 decided" count is consistent with the spoken "thirteen of those lines so far" and with the thirteen rows visible on the card, and nothing more. The underlying menu-bar app is not named and I found no repository for it. Treat all of these as unverified.

Key Takeaways

  • Verified: Michael Nygard published "Documenting Architecture Decisions" on 15 November 2011, and his template is Title, Status, Context, Decision, Consequences, with no date field.
  • Verified: adr-tools stamps an ISO 8601 Date: automatically via date=${ADR_DATE:-$(date +%Y-%m-%d)}, and its ADR 0008 of 2017-02-21 records the reason as "An important component of the history is the time at which a decision was made."
  • Verified: Anthropic's own documentation states "Each Claude Code session begins with a fresh context window," which supports the video's empty-context premise.
  • Correction: The auto-transcript's "Read only when the disk is full" is wrong. The on-screen text reads red only, disk 90% full, and 09-16 Red only at 90% disk.
  • Correction: The paper is "Programming as Theory Building," not "Programming is theory building." The card's "A program is the theory in its builders' heads" is a paraphrase. Naur wrote that the theory "could not conceivably be expressed, but is inextricably bound to human beings."
  • Partial correction: Naur's death condition is team dissolution, not a single builder leaving, and his conclusion that "program revival ... merely from the documentation, is strictly impossible" cuts against using a dated design doc as the remedy he would endorse.
  • Clarification: MADR 4.0.0 (17 September 2024) does include an optional date, but it is defined as "when the decision was last updated," which is not the same field the video is describing.
  • Context: Anthropic's memory page notes CLAUDE.md is "context, not enforced configuration," and points at a PreToolUse hook when a decision must actually be enforced.
  • Context: The video is small at capture, with 588 views and 25 likes about thirteen hours after posting, so engagement is not evidence the practice is widespread.

Resources

Published October 7, 2026. Writeup generated from a favorited TikTok.