> [!question] Lingering Questions > - What would change this claim if it turned out to be wrong, and where's the strongest counter-evidence? > [!idea] Article Ideas > - Because models are grown rather than hand-designed, the lever available to practitioners is the harness and the prompt around the model, not the model's weights — Anthropic cut ~80% of Claude Code's system prompt toward context over constraints. > [!info] Cross Reference > - See MOC links below for the cluster this note belongs to. > [!abstract] Vocabulary > - (no source-specific jargon requiring definition in this piece) --- ## Metadata - **Source:** [[3. Resources/Readwise/Articles/Field Guide to Fable — Thariq Shihipar, Anthropic]] - **Author:** Thariq Shihipar - **URL:** https://www.youtube.com/watch?v=9fubhllmsBU#transcript - **Type:** article - **Citekey:** fable-field-guide-shihipar ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: the harness-and-prompt leverage claim is this map's "where the human sits in an AI pipeline" thread, stated by an Anthropic insider - [[3. Resources/Permanent Notes/An Agent Prompt Must Name Its Stopping Point]] — extends: the field guide's harness-shaping advice needs the stopping rule that note supplies, or the agent runs past the task - [[3. Resources/Permanent Notes/Agentic Tooling Shifts the Developer From Author to Editor]] — evidence-for: Shihipar's account of leverage moving from writing code to shaping the run is the author-to-editor shift seen from the vendor side ## Author's thesis **Q1 (What is the author saying?):** As coding models get more capable, the human's leverage moves away from writing code and toward two things: shaping the harness/prompt the model runs in (since the model itself is "grown, not designed"), and specifying intent well enough to resolve the "unknowns" the agent surfaces when its map (the spec) doesn't match the territory (the actual codebase). **Q2 (Why should the reader care?):** It reframes a widely-assumed cause-and-effect relationship with concrete, checkable evidence, which changes what action is worth taking next. ## Key arguments 1. Because models are grown rather than hand-designed, the lever available to practitioners is the harness and the prompt around the model, not the model's weights — Anthropic cut ~80% of Claude Code's system prompt toward context over constraints. 2. The new bottleneck with capable agents is matching the human's mental map (spec/plan) to the actual codebase territory; gaps are named "unknowns" and surfacing them is a discipline in itself. 3. Building software is getting easier, but the article's honest counterweight is that generating actual value from what gets built is still hard — a caveat against pure build-velocity hype. ## Claim atoms ### Models are grown, not designed; the harness contains them > [!quote] ID:rw_fable-field-guide-shihipar_1 > the models are grown, not designed… what contains them is us, right? The harness we put them in and the way we prompt them - Timestamp: ~02:42 (Part 1, unhobbling). - Why this matters: reframes leverage toward improving prompts/harness, not the weights. ### Newer models want context, not constraints > [!quote] ID:rw_fable-field-guide-shihipar_2 > this new class of models want… a smaller system prompt. The examples tend to constrain it… we tried to give it context and not just constraints. - Timestamp: ~06:00 (removed ~80% of Claude Code's system prompt). - Why this matters: reverses prior prompt best-practice — fewer examples, more context. ### The map is not the territory; gaps are "unknowns" > [!quote] ID:rw_fable-field-guide-shihipar_3 > the plan and prompt and spec that I have in my mind is the map… the territory is the actual code base… whenever Claude runs into something in the territory that's not in the map, I call that an unknown - Timestamp: ~09:18 (Part 2, finding your unknowns). - Why this matters: names the new human bottleneck with capable agents. ### Building is easier, but generating value is still hard > [!quote] ID:rw_fable-field-guide-shihipar_4 > building is easier, but generating value is still hard - Timestamp: ~18:28 (Part 4, being unreasonable). - Why this matters: honest counterweight to build-velocity hype. ## Permanent candidates #### Agents Shift Human Work To Specifying Intent As a coding agent becomes able to traverse most of a problem on its own, the human's leverage shifts from writing the code to specifying intent well and resolving the unspecified decision points — the "unknowns" — that the agent surfaces along the way. ## Processing checklist - [x] Bib capsule complete - [x] Thesis captured (Q1 + Q2) - [x] Key arguments extracted (≤3) - [x] Permanent note candidate(s) spun out - [ ] Cross-links reviewed by next sweep