Add a learning path through the pattern collection - #51
Conversation
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
||
| ## The story in one paragraph | ||
|
|
||
| You've been handed a brilliant, absurdly well-read colleague who has **no memory**, **limited attention**, **rolls dice on every answer**, **never says no**, and **you can't see what it's thinking**. The whole collection is the story of compensating for those five facts, in escalating scope: first you manage its *memory*, then its *attention*, then you learn to *verify* everything and even exploit the dice, then you turn the one-way command channel into a *dialogue*, meet it in its native *medium*, replace fragile instructions with *mechanisms*, and finally scale from one session to *systems of agents that improve themselves*. Every obstacle is one of the five facts showing up; every anti-pattern is a natural-but-wrong reaction to it; every pattern is the working counter-move. |
There was a problem hiding this comment.
Fix: "you can't see what they're thinking" (and other instances of "it's" in this paragraph.)
|
|
||
| You've been handed a brilliant, absurdly well-read colleague who has **no memory**, **limited attention**, **rolls dice on every answer**, **never says no**, and **you can't see what it's thinking**. The whole collection is the story of compensating for those five facts, in escalating scope: first you manage its *memory*, then its *attention*, then you learn to *verify* everything and even exploit the dice, then you turn the one-way command channel into a *dialogue*, meet it in its native *medium*, replace fragile instructions with *mechanisms*, and finally scale from one session to *systems of agents that improve themselves*. Every obstacle is one of the five facts showing up; every anti-pattern is a natural-but-wrong reaction to it; every pattern is the working counter-move. | ||
|
|
||
| **Legend:** ⛰ obstacle (the terrain — can't be removed) · ⚠️ anti-pattern (the wrong turn) · 🧩 pattern (the move) · ↳ special case / composition marker |
There was a problem hiding this comment.
I'd perhaps change "the wrong turn" to "the mis-step", and "the move" to "the trodden path"?
|
|
||
| - ⛰ Limited Focus — attention dilutes or fixates | ||
| - ⚠️ Distracted Agent — one agent for everything; rules silently ignored | ||
| - ⚠️ Obsess Over Rules — piling on rules makes each one weaker |
There was a problem hiding this comment.
Refers forward to "Refinement Loop" pattern
| *It fabricates confidently and declares victory early. Never let output go unverified — and make verification cheap enough that you don't skip it.* | ||
|
|
||
| - ⛰ Hallucinations — it invents APIs (mercifully self-revealing in code) | ||
| - ⚠️ Perfect Recall Fallacy — expecting it to remember library details from training |
There was a problem hiding this comment.
Forward reference to "Playgrounds" pattern. (Could be mitigated if this pattern talked about letting LLMs/Agents "Play" which then sets up "Playgrounds" perfectly...)
| - ⛰ Solution Fixation — first plausible hypothesis becomes "the answer" | ||
| - 🧩 **Feedback Loop** — give it a success signal and permission to iterate until green | ||
| - ⛰ Keeping Up — it generates faster than you can review | ||
| - ⚠️ Flying Blind — the surrender: skip review, vibe-code, understand nothing |
There was a problem hiding this comment.
Forward reference to "Active Partner" Pattern
| - ⛰ Degrades Under Complexity — errors compound with step size and artifact size | ||
| - 🧩 **Chain of Small Steps** — small, verified, committed steps; complexity handled by you, not it | ||
| - 🧩 **Canary in the Code Mine** — flip the signal: when the AI struggles, your *codebase* is telling you it needs refactoring | ||
| - 🧩 **Surface Change Attractors** — the same signal, mined from history: hotspots and change coupling become the agent's redesign constraints |
There was a problem hiding this comment.
This pattern isn't in the catalogue at https://lexler.github.io/augmented-coding-patterns/pattern-catalog/
There was a problem hiding this comment.
Yes, because it was still an unmerged pull request. I merged it now
There was a problem hiding this comment.
Mice pattern. Reminds me of Vlad Khononov's book on coupling that I still need to read...
| - 🧩 **Yak Shave Delegation** — ↳ Background Agent for the tangent that's blocking you | ||
| - 🧩 **Orchestrator** — the integration agent: monitors, merges, resolves, keeps trunk green | ||
| - 🧩 **Smart Plan, Cheap Execution** — tier the models: the strongest plans, a cheaper one executes the well-specified plan | ||
| - 🧩 **Escalate When Stuck** — the reverse tiering: cheap by default, consult the strong model on defined triggers |
There was a problem hiding this comment.
This one isn't in my catalogue - is it awaiting merging?
There was a problem hiding this comment.
Renamed to follow Anthropics naming: https://ai-coding-patterns.dev/patterns/advisor-strategy/
| ### Chapter 11: Many Hands | ||
| *One focused agent is good. Several, each focused, are better — if someone integrates.* | ||
|
|
||
| - 🧩 **Chunking** — orchestrator stays strategic, subagents execute — like humans delegating practiced skills |
There was a problem hiding this comment.
Should we have a "Subagent" pattern? Even an "Agent" pattern...?
There was a problem hiding this comment.
This pattern also refers forward to the "Orchestrator" pattern
2 participants
A narrative organization of the whole collection into 12 chapters, plus the duplicate/special-case analysis behind it. Extends Lada's 27-station semantic map to every document.
Two things to decide before this leaves draft:
learning_path.htmlis a rendered view of the same content. Nothing else in the repo ships a standalone HTML file, so it may not belong here — happy to drop it and keep only the markdown.learning_path_map.svgis not included: it disappeared from the working tree before I could commit it.