ICM Is a Real Published Method, and This 58-Second Explainer Gets Its Five Layers Right

The folder layout animated in this video matches a March 2026 arXiv paper almost box for box, and the paper's own authors admit they have no controlled evidence that it works better than one long chat. The clip runs 58 seconds and contains no face, no screen recording, and no code editor. It is a sequence of animated slides on a near-black background, styled like a terminal readout: monospace type, orange and cyan and green outlines, a header reading AGENT SYSTEMS / ICM, a run label run_017 that persists across every slide, and a footer reading ICM · SEQUENTIAL WORKFLOW EXAMPLE. A progress bar fills along the bottom. A voiceover narrates over the slides, and the visible text is slide content rather than burned-in captions. The argument is that a single agent loses context between sessions the same way multiple agents lose it across handoffs, and that a numbered folder structure with a per-stage contract file fixes the handoff rather than the memory. At the time I pulled the metadata the video had 1,655 views, 67 likes, and zero comments.

ICM is not the creator's coinage

The description asserts that ICM stands for Interpretable Context Methodology. That checks out against a primary source. Interpretable Context Methodology: Folder Structure as Agentic Architecture was submitted to arXiv by Jake Van Clief and David McDermott on 17 March 2026, with a v2 the following day. The listing describes 28 pages, 5 figures, 2 tables, and 54 references, and gives a contact address at Eduba alongside a University of Edinburgh affiliation. The paper states its thesis plainly: it "replaces framework-level orchestration with filesystem structure."

There is a reference implementation too. The Interpretable-Context-Methodology repository publishes the workspace layout under the MIT license, and a companion repository, icm-architect, packages the method as a Claude Code skill that scaffolds a new workspace or restructures an existing folder into one. A third-party variant, icm-template, ships the same idea as a drop-in template. So the video is describing something with a paper, a spec, and at least three public implementations behind it, which is more than most TikTok methodology acronyms can claim.

The slides render the paper's layer model, with renamed folders

The central slide, labeled ICM / FILESYSTEM ROUTE, draws a tree. At the root sits workspace/, tagged WORKSPACE ROOT, with an arrow to a box containing CLAUDE.md and AGENTS.md labeled L0 · IDENTITY. Below the root is CONTEXT.md tagged L1 · ROUTER, and beside it _config/ tagged L3 · REFERENCES. CONTEXT.md branches into three stage folders, 01_research/, 02_script/, and 03_production/, each tagged L2 · STAGE. Each stage then produces one artifact on an L4 · OUTPUT row: research.md, script.md, and render.mp4. A caption reads ORDER IN NAMES · STATE IN FILES.

That maps to the paper's five-layer hierarchy exactly. The paper defines Layer 0 as global identity ("Where am I?"), Layer 1 as workspace routing ("Where do I go?"), Layer 2 as the stage contract ("What do I do?"), Layer 3 as persistent reference material such as design systems and voice guides, and Layer 4 as per-run working artifacts. The video's layer tags and the paper's layer definitions agree on all five.

The folder names differ cosmetically. The published repository nests stages one level deeper and hyphenates them, as stages/01-research/ and stages/02-script/, with output/ directories inside each stage acting as handoff points. The video flattens that to 01_research/ at the workspace root with underscores. Either form preserves the property the caption names, which is that execution order lives in the filenames. I would treat the video's tree as a simplification for a vertical slide rather than a competing spec.

The stage contract slide is the strongest part

Four consecutive slides animate 02_script / CONTEXT.md as a three-column contract: INPUTS, PROCESS, OUTPUTS. The columns fill in sequence. First research.md arrives tagged WORKING and voice.md arrives tagged STABLE. The process column moves from WAITING to AGENT WRITES to WRITTEN. Then script.md appears in the outputs column stamped run_017. A HUMAN REVIEW GATE box below shows AWAITING script.md, then flips to APPROVED · run_017, at which point an arrow releases into 03_production/, which had been sitting in a WAITING FOR REVIEW state.

The repository confirms this three-part contract structure: each stage's CONTEXT.md specifies inputs with justification, a step-by-step process, and the artifacts produced. It also confirms the review behavior the animation shows, describing the intermediate output of each stage as a file a human can open, read, edit, and save before the next stage runs, with edits persisting into the following stage. The distinction the slide draws between research.md as WORKING and voice.md as STABLE is the paper's Layer 3 and Layer 4 split, where reference material is meant to be internalized as constraint and working artifacts are meant to be processed as input.

Layer 0 is where the video adds something the paper does not

One slide, CODING AGENT DEFAULTS, puts Claude Code and Codex side by side, each with a persistent instruction file feeding a model session: CLAUDE.md for Claude Code, AGENTS.md for Codex. Its subtitle reads PROJECT GUIDANCE ≠ ENFORCEMENT, and a box at the bottom reads READS SELECTED BY AGENT · WORKFLOW GUIDANCE IS POSSIBLE.

The enforcement point is correct and verifiable. Anthropic's own Claude Code memory documentation states that instruction files are loaded at the start of every conversation and that "Claude treats them as context, not enforced configuration," pointing users to a PreToolUse hook when they need to block an action outright. AGENTS.md is a real cross-vendor format, now stewarded by the Agentic AI Foundation under the Linux Foundation, supported by more than 20 coding agents and claimed on its homepage to appear in over 60,000 open-source projects.

The pairing is slightly tidier than reality in two ways. First, the same Anthropic page says Claude Code can read a repository's AGENTS.md files, either on their own or alongside CLAUDE.md, so the clean Claude-Code-versus-Codex split the slide draws is a presentational convenience. Second, the ICM paper names CLAUDE.md as its Layer 0 identity file and does not discuss AGENTS.md at all, while the community icm-template repository uses a separate IDENTITY.md for that role. The video's decision to show both coding-agent files as Layer 0 is a reasonable generalization, and it is the video's own, not the paper's.

The context argument rests on borrowed evidence

The narration opens by describing how "instructions corrections and old outputs compete for the same context" in long chats. That premise has independent support. Lost in the Middle found that model performance is highest when relevant information sits at the beginning or end of the input and degrades significantly when the model must retrieve from the middle. Chroma's Context Rot report, published 14 July 2025, tested 18 models including GPT-4.1, Claude 4, Gemini 2.5, and Qwen3, and found that performance varies significantly with input length even on simple tasks.

The paper's own numbers deserve a label. It states that the total context delivered at any given stage "typically ranges from 2,000 to 8,000 tokens," against a monolithic approach that "can easily reach 30,000 to 50,000 tokens." Those read as practitioner estimates rather than instrumented measurements, and the paper does not present a benchmark comparing the two. The authors are candid about this. Their empirical base is three production workspaces, an invite-only community of 52 members, and intervention-pattern observations from 33 users gathered through conversation rather than structured interviews or logs. They state there is no controlled comparison between staged context loading and monolithic prompting, and that testing was confined to Claude Opus 4.6 and Sonnet 4.6. So the mechanism is plausible and the supporting long-context research is real, while the specific claim that ICM improves output quality has not been measured.

The closing caveat is the paper's caveat

The video ends by conceding scope: "code still enforces parallelism retries permissions and gates, ICM makes the checkpoints visible." The final slide, WHO SELECTS CONTEXT?, compares one chat with a mixed window, an agent default reading guidance on its own, and ICM's stage contract with file handoff, then lists BRANCHING, PARALLELISM, RETRIES, and PERMISSIONS under the header FOLDERS DO NOT ENFORCE.

Every item on that list appears in the paper's own limitations. It describes ICM as sequential by design, handles failure by manual re-run of the failed stage, leaves permissions unaddressed in a local-first workspace, and routes conditional branching through a human decision between stages rather than automated logic. The authors frame these as deliberate trades, preferring simplicity of implementation over feature completeness. A 58-second explainer that reproduces its source's limitations section is doing better than most.

Key Takeaways

• ICM is a documented methodology with an arXiv paper, an MIT-licensed reference repository, and a Claude Code skill implementation, rather than a term invented for the video.
• The five layers shown on screen, from CLAUDE.md at L0 through stage outputs at L4, match the paper's hierarchy. The folder names on screen are flattened compared to the repository, which nests stages under stages/ and uses hyphens.
• The per-stage CONTEXT.md contract of inputs, process, and outputs, plus the human review gate between stages, are both faithful to the published spec.
• The "2,000 to 8,000 tokens per stage versus 30,000 to 50,000 monolithic" figures come from the paper as estimates. No benchmark backs them, and the authors say no controlled comparison exists.
• Long-context degradation is independently supported by Liu et al. and by Chroma's 18-model report. The specific benefit of ICM's staging is inferred from that work rather than demonstrated.
• Transcript accuracy: the Whisper output is lowercase and unpunctuated but tracks the audio closely, including "agents.md and claude.md." One clause, "icm does numbered folders separate for the work," reads garbled and should be understood against the slide that shows 01_research/, 02_script/, and 03_production/ under the caption ORDER IN NAMES · STATE IN FILES. The video carries no burned-in caption track, so on-screen text is slide content and does not duplicate the narration.
• Unverified: the exact run_017 example workflow, the three production workspaces, and any before-and-after quality difference attributable to ICM. The slides are illustrations of a structure, not a recorded run. I also could not confirm any connection between the poster and the paper's authors.

Resources

• Interpretable Context Methodology: Folder Structure as Agentic Architecture -- The arXiv paper by Jake Van Clief and David McDermott that defines the five-layer model, the stage contract, and the stated limitations the video repeats.
• RinDig/Interpretable-Context-Methodology -- MIT-licensed reference workspace showing the canonical folder layout and the inputs/process/outputs contract file.
• RinDig/icm-architect -- Claude Code skill that scaffolds a new ICM workspace or restructures an existing folder into one.
• ktnCodes/icm-template -- Third-party template using IDENTITY.md plus CONTEXT.md at the root, useful for seeing where implementations diverge on Layer 0.
• Claude Code memory documentation -- Anthropic's page confirming instruction files load every session and are treated as context rather than enforced configuration.
• AGENTS.md -- The cross-vendor instruction file format stewarded by the Agentic AI Foundation, supported by 20-plus coding agents.
• Lost in the Middle: How Language Models Use Long Contexts -- The positional-degradation result the ICM paper cites as its justification for scoped context loading.
• Context Rot: How Increasing Input Tokens Impacts LLM Performance -- Chroma's July 2025 evaluation across 18 models showing performance varies with input length on simple tasks.