Change log

DocsRe-audit every model and lesson against Tide's latest approach

Checked all 10 mental models and 17 lessons against Tide's current approach (tide main a3ed538: evals/LIFECYCLE.md, docs/EVALS_STRATEGY.md §2.10, §2.15, §7): real Coach on real data, traced; owner grades first; failures grouped by first upstream cause; fixes proven by frozen replays or live re-runs; code checks; a fresh round 2.

DocsLesson 3 follows the steps Tide actually took; red lines are live

Checked dimensions-vs-slices against tide main (a3ed538): docs/EVALS_STRATEGY.md §2.1-2.6, evals/taxonomy.ts, evals/scenarios/build_round1.py + round-1.json, evals/labels/round-1.json, evals/LIFECYCLE.md Phase 1.

NewLeaner lessons on the Evals page, same as the mental models

Each of the 17 lessons on the Evals page now shows its TL;DR, the idea, its real example and "Mental model / why it matters". Dropped from the page, as for the mental models in #83: "How to apply", "Provenance & caveats", "Connections". They stay in the notes (concept graph + fact-check record); a one-line link under each lesson opens the full note in the reader.

NewLeaner mental models on the Evals page

On the Evals page, each mental model's "go deeper" view now shows just the idea ("The idea, simply" / metrics / principles) and "Mental model / why it matters". Dropped from the page: "How to apply (in practice / consulting)", "Provenance & caveats" and "Connections".

FixCircles are only for loop steps; models and lessons always say M#/L#

The Evals page had bare numbers in the same visual style for two different things: green circles "1" (foundation) and "10" (across it all) were mental models M1/M10 but read as steps 1 and 10, next to the loop's step circles 1-8. Model cards repeated it (a green "2" for M2).

NewOne structure, one numbering — the page walks the loop the picture draws

Two pieces of feedback on the Evals page: 1. Not every lesson and mental model was reachable from the loop diagram. 2. The diagram's numbering and structure didn't match the content below. There were three parallel schemes: steps 1-8, models M1-10 in their own order, and lessons L1-17 under five different stage headings.

DocsDropped the "On the concept graph" section; bring Tide-state claims up to date

Evals page: removed the "On the concept graph" section, its "4 · On the concept graph" jump link, the script that filled it, and its CSS. The concept nodes stay on the concept graph itself. - Audit of everything else the page renders (28 notes, the loop, footer): five caveats described Tide's state as of September. Updated against tide docs/EVALS_STRATEGY.md at #341: - dimensions-vs-slices, sort-before-you-count, whose-fault-is-the-failure: the scripted suite's checks, answer keys, criteria and judge WERE repaired (re-baseline 2026-09-21: 66-71 of 92), then the suite was archived in October for the rebuilt 100-scenario real-data suite; the gate was never moved onto red lines on it. Certifying the rebuilt suite's per-type judges is on tide's backlog (F4.6-F4.12). - whose-fault-is-the-failure: "has since grown to 92" -> grew to 92, then archived. - dont-copy-derive: the "planned" generated test list is now done (generate_evals_archive.ts builds docs/evals-archive.html, tide #339). - Checked and still current: both external links (200), the temperature deprecation claim (re-verified on platform.claude.com), and the flywheel script the offline-evals note cites (still on tide main).

NewAdded the eval lifecycle + five fix-and-prove lessons; picture shows the full loop

Lands tide's evals/lessons/kc-00..05 (tide #334), which the tide session could not write here.

NewRebuild the Evals page as a readable guide — ten mental models

Vik's feedback on the hub: too many clicks (every idea opened a modal), the top didn't flow into the lessons, and his personal question list shouldn't be on the public site. The page is now something you read top to bottom:

NewAdded "your 10 questions" — the evals mental model from Hamel, Shreya & co.

Vik's 10 eval questions, each answered as one atomic note distilled from: Hamel Husain & Shreya Shankar's evals FAQ (+ individual FAQ pages on RAG, agents, multi-step workflows, guardrails vs evaluators, CI vs production, trusting an automated eval, and trace sampling); their two Lenny's Newsletter pieces; Hamel's "Your AI Product Needs Evals"; Saha & Husain's "Do Automated Evals Work?" (Parlance Labs); and Aman Khan's PM guide on Lenny's.

NewAdded lesson 12 — whose fault is the failure

The last unlanded draft from the tide packet (another session copied all 12 drafts to tmp/evals-lessons/ before its worktree was deleted; drafts 2-11 were byte-identical to what's already landed, 12 was new).

FixDon't let a backslash escape in the diff crash the build

The "Update change log" workflow failed on #74 with:

NewAdded lessons 10-11, correct lesson 9, add the D1-D5 legend

Swept ai_knowledgecenter for work other sessions left unlanded, and landed it.

DocsDropped the index/count line from the Evals hub hero

Removes "Want the running index? … · 9 lessons so far" from the hero of docs/concepts/evals.html, along with the one script line that filled its lesson count (it would throw without its span and stop the concept cards rendering).

NewAdded the Evals section: 8 lessons, LLM-as-judge concept, hub page

Lands the Tide evals-lessons packet (tide PR #234) as a new Evals section.

Added offline-vs-online evals (the agent is real, the world is frozen)

New foundations note distilled from rebuilding the Tide Coach eval harness (tide PR #234): offline evals as a controlled experiment, the offline/online pair and which graders work where, the trace-to-fixture flywheel, and two lived failure modes (unfrozen clock; dead harness scoring a fake 0%).

DocsAdded PM AI on-ramp learning plan (unpublished)

Two-pager + appendices distilling the knowledge base into a learning path for a senior PM new to hands-on AI. Framework: prompt -> context -> harness -> loop, plus a 6-week Claude-ecosystem path (Claude.ai -> Cowork -> Claude Code) and a cross-vendor cheat sheet.

DocsClose harness-vs-context migration — Tide source page removed

The harness-vs-context content was distilled into this repo on 2026-06-02 (canonical note + the 12-component anatomy note + the interactive mental-model render). The Tide-side removal half of that migration was never executed, leaving the source page live on the Tide portfolio site.

DocsAdded workflow-journal-resume — durable orchestration via resumeFromRunId

New atomic note: workflow runs journal every agent() result keyed by (prompt, opts); resumeFromRunId replays the unchanged prefix from cache and re-runs only failed/edited calls. Durability — not parallelism — is the decisive reason to script destructive multi-surface work as one workflow instead of N ad-hoc agents.

FixUncap home hero h1 + sub so both span full content width

docs/index.html had home-only inline overrides (.hero h1 max-width:16ch, .hero .sub max-width:64ch) that re-capped the hero after the shared stylesheet was widened. Set both to max-width:none so the heading and subtext run to the container edge (left-aligned), matching the other hubs.

DocsDeclutter mental-model page; unify all four control levels into one template

The page was lopsided and dense: only Loop (④) had a full section, while Prompt/Context/Harness (①②③) got a one-line step card each, on top of two always-open accordions and a large combined callout.

FixLet hero subtext span full content width (left-aligned)

The hero `.sub` was capped at max-width:62ch, so the description wrapped early and left a large empty gutter on the right. Drop the cap to max-width:none so it fills the .wrap container (1140px), still left-aligned. Bump the stylesheet cache-buster so the change ships to browsers.

DocsAdded where×trigger matrix + containment-vs-discrimination lens to loop axis

Ground the mental-model loop section in the official docs. The #loop section explained what loop engineering IS but not how to PICK among /loop, /goal, hooks, routines & cron — the question the docs actually organize around.

DocsDe-dup Decide vs Comparison — split by intent (choose vs compare)

The two pages cover the same ten primitives but on different axes: Decide is symptom-first (need → tool), Comparison is tool-first (tool → properties). Keep them separate, remove the genuine overlaps.

DocsCross-link primitives page to terminology fixes

Add a one-line signpost in the primitives.html hero pointing to the matrix's Terminology fixes section (matrix.html#terms). Closes the discoverability gap for readers confused by primitive names without moving the disambiguation content off the comparison page where it belongs.

DocsLift breadth-vs-depth to mental-model as Axis 3, leave scored table on matrix

Breadth vs depth is a reasoning axis, not lookup data — same genre as Soft-vs-Hard and Who-holds-the-plan. Move the frame + rule of thumb to the mental-model page as Axis 3 (with a note on how it differs from Axis 2), keep the per-primitive scored bars on the comparison matrix, and cross-link both ways so the concept lives once.

DocsUpdated Mental model

DocsRecord changelog deep-link anchor-validation lesson

Capture the non-obvious gotcha: nearest-anchor is computed from the file at the commit (for line alignment) but links to the live page, so anchors must be validated against the current working-tree file or they 404.

NewLinked changelog to live pages + show exact text that changed

The change log's "What changed" list linked every file to its GitHub commit diff — raw markup that doesn't tell you what a page now says or where to look.

DocsDisambiguate /loop, /goal, routines, cron & the tool-call family

Adds connected teaching material for two confusable clusters, seeded by Matt Van Horn's "WTF Is a Loop?" article and verified against official docs.

FixLet the comparison matrix scroll horizontally

The matrix table sat in the 1fr column of `.layout`. Grid items default to min-width:auto, so the wide table ballooned the column past 1fr instead of letting `.matrix-scroll`'s overflow:auto take over — and body's overflow-x:hidden then clipped it, leaving no way to reach the right edge.

NewAdded "loop engineering" + add source-of-truth verification gate

Land Addy Osmani's loop-engineering thread as a new patterns-layer concept (loop-engineering) + source note (designing-loops-not-prompts), wired into the spine, Home Note, and concept graph (0 dropped edges, no orphans).

NewAdded active Overview tab to top nav

Landing page now shows an "Overview" button in the nav, auto-selected with the active (you-are-here) treatment. Existing hero + explore band content is wrapped in an #overview section.

FixGave cross-view pill its own color, distinct from the active tab

The top-right "Exam Prep" / "Concepts" pill shared the exact green→blue gradient + dark text used for the active nav tab, so it read as "you are here" rather than "jump to the other view". Split the pill/CTA off into its own rule with a purple→blue gradient + white text so the two affordances are visually distinct. Applied to both stylesheets (concepts + home) and bumped the CSS cache-bust version across all pages.

NewRebuild landing page in the portfolio / Product Cockpit style

Align the AI Knowledge Center landing page to the portfolio (vikkybharadwaj.github.io) + Product Cockpit landing experience: eyebrow -> gradient headline -> lede -> CTA buttons, and a 2-up glassy accent-card "Two ways in" band (Concepts / CCA-F Exam Prep) with accentbar, status badge, icon, tagline, and tag chips. Reuses the portfolio-aligned design tokens already in style.css.

DocsUpdated home page

DocsUpdated home page

DocsUpdated home page

StyleAlign Knowledge Center UI to the portfolio / Product Cockpit design

Re-skin the whole site to the same visual language as vikkybharadwaj.github.io and the Product Cockpit: JetBrains Mono throughout, dark-only #0a0f1a canvas, glassy translucent cards, ambient glow blobs, and the signature green→blue gradient on nav pills, card accents, brand dot, and primary buttons.

NewDefine edge types, fact-check vs docs, CLI scope

Three follow-ups on the dependency map:

NewReframed the graph as a tech-stack dependency map

The concept graph plotted notes (distilled articles) wired by note relationships — it taught the reading list, not the system. Reframe it to plot the tech stack itself.

DocsAdded Learn Harness Engineering curriculum as 4 connected notes

Distill the 12-lecture "Learn Harness Engineering" site into 4 atomic notes mapped to the harness's underlying subsystems rather than one-per-lecture:

NewRedesigned concept graph for legibility (SVG chips + spine lanes)

Replaces the canvas force-blob (bare circles with overlapping full-length titles) with a deterministic SVG concept map:

DocsRecord Pages-breaking orphan-gitlink learning

FixRemoved orphan worktree gitlink that broke GitHub Pages

`.claude/worktrees/changelog-home` was accidentally committed as a gitlink (mode 160000) with no .gitmodules entry. GitHub Pages' checkout runs `git submodule update --init` and fails with "No url found for submodule path", so every pages-build-deployment has been failing and the live site was frozen on a stale build. Untrack it and gitignore .claude/worktrees/ so it can't recur.

NewRestyle change log like the Claude Code changelog

DocsAdded plain-English git glossary + link from Home Note

Reference doc defining the common git/GitHub terms used to maintain this base (commit, branch, merge, push, pull, PR, worktree, and more). Lives outside knowledge/notes/ since it isn't a Claude-specific atomic note, so it stays out of the concept graph and the note-validation hook.

DocsPublished site and tooling updates

ChoreAdded workflow_dispatch to change-log workflow

Allows manual runs/backfills from the Actions tab. Also serves as the first post-registration push to main, which triggers the workflow (a newly-added workflow does not run on the same push that introduces it).

DocsPublished site and tooling updates

NewAdded auto-generated change log + move home tagline under title

Home: move the "living, personal learning hub …" credit up as subtext under the main title; footer now carries a "Change log →" button below the line. - New docs/changelog.html: native, same-site page listing every shipped PR newest-first — date·time, title, description, and links to the PR + exact diff. - scripts/build-changelog.py regenerates the log from git history (squash-merge "(#NN)" commits), idempotently rewriting between CHANGELOG markers. - .github/workflows/changelog.yml runs it on every push to main so each merged PR auto-appears; self-commits with [skip ci] to avoid loops.

DocsAdded routines primitive and Claude code routines note; updated 7 pages

DocsAdded worktree base ref fresh vs head note; updated concept graph and big-picture map

DocsAdded dynamic-workflows notes + Concepts card lands on Overview

docs: land dynamic-workflows article as two connected notes

FixHub breadcrumb says "Concepts", not "Primitives"

The #28 merge auto-applied #27's index.html breadcrumb edit (Concepts→Primitives) to the new Concepts hub, so the hub's crumb disagreed with its active rail item ("Overview"). Restore "Concepts".

NewGave Concepts a real landing hub; ship top-nav refactor

The "Concepts" tab (and every breadcrumb "Concepts") landed on the Primitives page because concepts/index.html *was* the Primitives page. Add a dedicated Concepts hub at index.html (overview + "ways in" cards, mirroring the exam hub) and move the primitives field guide to its own concepts/primitives.html. Add a uniform "Overview" rail item across all concept pages and retarget every Primitives rail/breadcrumb link.

DocsTidied up breadcrumbs and primitive h1s

Drop the em-dash/parenthetical subtitles from primitive page <h1>s (MCP, Conventions, Policies, Skills, Dynamic workflows) and the mental-model deep-dive link, and fix the Concepts/About-the-exam breadcrumbs so each page's crumb matches its left-nav label.

DocsRenamed Deep dives→Primitives + land Concepts card on Mental model

docs: rename "Deep dives" concept to "Primitives"

DocsKept the home "AI Knowledge Center" h1 on one line

The shared .hero h1 rule caps width at 14ch, which wrapped the 19-char home title onto two lines. Override it in the home-only style block: drop the width cap, force nowrap, and use a clamp() font-size so it still scales down on narrow phones instead of overflowing.

DocsRenamed "Deep dives" concept to "Primitives"

Relabel the Concepts left-nav item "Deep dives" → "Primitives" across all 15 concept pages (it is the primitives index), and update the matching descriptive mentions in the hub hero, README, and CLAUDE.md. Left genuinely distinct uses untouched (exam-prep "deep dives", the CSS accordion comment, and the inline "dynamic-workflows deep dive" link).

DocsReordered Concepts rail — Mental model first

New rail order: Mental model → Deep dives → Concept graph → Comparison matrix → Decide (then the divider + Legend). Leads with the mental model as the orientation entry point. Applied across all 15 Concepts pages; active states and relative paths preserved.

DocsAdded Legend as a left-nav tab below Decide

Surface the badge legend as its own rail tab across every Concepts page, separated from the primary nav by a thin divider so it reads as a cross-reference rather than a peer of the main views. Adds a .rail-sep divider style and marks the tab active on legend.html.

DocsMade Deep dives a left-nav tab, drop the Overview sheet

Rename the "Overview" rail tab to "Deep dives" across every Concepts page and refocus the landing page (index.html) on the nine primitive deep-dive cards. Removes the "two ways in" overview framing and the redundant "Go broad" concept-graph pointer section — the concept graph is already its own nav tab. The #primitives anchor is preserved so the primitive-page breadcrumbs still resolve.

DocsMoved breadcrumb to a full-width horizontal heading bar above the rail

Moves the breadcrumb out of the left rail into its **own full-width horizontal heading bar** at the top of every page, consistent across all 18 Concepts/Exam pages.

FixFixed the left rail — cache-bust style.css + add rail rules to root stylesheet

Fixes the left rail rendering unstyled (stacked at the top instead of a left column).

DocsLeft-column nav rail across all pages + exam as a peer of Concepts

Replaces the horizontal top nav bar with a consistent **left-column rail** on all 18 Concepts/Exam pages, and fixes the section hierarchy.

DocsAdded an accessible badge legend + link it from every pill-row

Adds an accessible **Legend** for the badges that appear on every primitive page (e.g. on Policies: `Governance` · `Hard · client-enforced` · `GA`).

DocsSimplified landing page to one title + two cards

Strips the landing page down to **one title + two cards**, per request.

DocsReframed Concepts overview as a clear broad-vs-deep structure

Makes the Concepts overview read as one intentional structure — **two ways in**:

DocsBreadcrumb nav + brand-home + drop Home/theme toggle

Site-wide navigation IA so every page shows where it sits and how to get back:

DocsRemoved theme toggle from the 9 primitives sub-pages

docs(concepts): trim Overview + drop theme toggle and nav-active highlight

DocsTrimmed Overview + drop theme toggle and nav-active highlight

Remove the header light/dark theme toggle from all 8 concepts pages (graph.html's inline toggle script already null-guards, layout/tag controls that reuse .theme-toggle styling are kept) - Drop the current-page nav-active highlight (.topbar nav a.active) so 'Overview' renders identically to the other nav links - Overview (index.html): remove the 'Start here / one idea' and 'The progression / ladder' sections; keep hero, the nine deep-dive cards, and the concept-graph pointer - Comparison matrix (matrix.html): remove the 'Terminology corrections' section

DocsAdded "Building & advising on AI stacks" consulting cluster

Answers the recurring question — "Claude Code is itself a harness, so when do I use it vs build my own, and how do I bootstrap a company's AI stack?" — and fills the two spine gaps the Big Picture flagged as empty (api, agent-sdk) plus lights up the products layer with its first scenario.

FixRoot redirect to /docs so Pages serves the HTML hub

When Pages serves from the repo root, it renders README.md instead of the docs/ landing hub. This root redirect (index.html takes precedence over README.md) bounces the site root into ./docs/, so the HTML experience loads regardless of whether Pages is set to root or /docs.

AI Knowledge Center landing hub + GitHub Pages

Add a hosted landing hub (docs/index.html) driven by a module registry, with the Claude Code Architecture module live, a self-contained design system, .nojekyll, and README/Pages docs.

Claude Code architecture field guide + CCA-F exam prep

Interactive multi-page learning site (mental model, comparison matrix, decision ladder, 9 primitive deep-dives, 6-scenario walkthroughs), a 136-question CCA-F practice bank with quiz engine, and cross-cutting knowledge incl. the canonical harness-vs-context note. Grounded in code.claude.com/docs and the official exam guide.