From 7d9ec98523dd76fcfbacc91548cb0b177207d76b Mon Sep 17 00:00:00 2001 From: Gregor Riegler Date: Wed, 15 Apr 2026 21:54:28 +0200 Subject: [PATCH 1/3] Add Poor Person's Skill pattern Co-Authored-By: Claude Opus 4.6 (1M context) --- documents/patterns/poor-persons-skill.md | 35 ++++++++++++++++++++++++ 1 file changed, 35 insertions(+) create mode 100644 documents/patterns/poor-persons-skill.md diff --git a/documents/patterns/poor-persons-skill.md b/documents/patterns/poor-persons-skill.md new file mode 100644 index 0000000..01a3166 --- /dev/null +++ b/documents/patterns/poor-persons-skill.md @@ -0,0 +1,35 @@ +--- +authors: [gregor_riegler] +related_patterns: + - semantic-anchors + - extract-knowledge +related_obstacles: + - limited-context-window + - hallucinations +--- + +# Poor Person's Skill + +## Problem +You're about to work on a topic the agent has plenty of training knowledge about, but none of it is active in the current context. And you don't have a skill or knowledge document at hand. Without priming, the agent produces generic output and you only discover the gaps by correcting them mid-task. + +## Pattern +Before giving the real instruction, ask the agent an open question: *"What do you already know about X?"* + +The answer itself is secondary. The point is that the agent's summary now sits in the context window, shaping everything that follows. You get most of the value of a dedicated skill file at the cost of one question — and you can see what the agent *thinks* it knows, so you can correct or supplement before the work starts. + +Use it when: +- The topic is well-represented in training data (frameworks, methodologies, known libraries) +- The task is too small to justify a written skill or reference doc +- You want a sanity check on the agent's baseline understanding before committing to an approach + +## Example +Before asking the agent to break a feature into a first slice: + +> What do I have to care about when doing vertical slicing? + +Or: + +> What is important for doing vertical slicing? + +The agent surfaces the usual concerns: keeping slices end-to-end, avoiding horizontal layers, picking the thinnest valuable path, deferring edge cases. Now when you ask it to propose a first slice, it draws on that surfaced knowledge instead of guessing. If the recall is wrong or thin, you've caught it before it shaped the work. From ff33bd4007f516ce7469216ea02e27642b721489 Mon Sep 17 00:00:00 2001 From: Gregor Riegler Date: Wed, 15 Apr 2026 21:56:40 +0200 Subject: [PATCH 2/3] Reword 'use it when' bullet to match Problem section Co-Authored-By: Claude Opus 4.6 (1M context) --- documents/patterns/poor-persons-skill.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/documents/patterns/poor-persons-skill.md b/documents/patterns/poor-persons-skill.md index 01a3166..081c30a 100644 --- a/documents/patterns/poor-persons-skill.md +++ b/documents/patterns/poor-persons-skill.md @@ -20,7 +20,7 @@ The answer itself is secondary. The point is that the agent's summary now sits i Use it when: - The topic is well-represented in training data (frameworks, methodologies, known libraries) -- The task is too small to justify a written skill or reference doc +- You don't have a skill or knowledge document at hand - You want a sanity check on the agent's baseline understanding before committing to an approach ## Example From 3c140273266c65237a29b81c8cf8b22241ad813b Mon Sep 17 00:00:00 2001 From: Gregor Riegler Date: Wed, 15 Apr 2026 21:57:38 +0200 Subject: [PATCH 3/3] Keep a single example question Co-Authored-By: Claude Opus 4.6 (1M context) --- documents/patterns/poor-persons-skill.md | 4 ---- 1 file changed, 4 deletions(-) diff --git a/documents/patterns/poor-persons-skill.md b/documents/patterns/poor-persons-skill.md index 081c30a..291f4c5 100644 --- a/documents/patterns/poor-persons-skill.md +++ b/documents/patterns/poor-persons-skill.md @@ -26,10 +26,6 @@ Use it when: ## Example Before asking the agent to break a feature into a first slice: -> What do I have to care about when doing vertical slicing? - -Or: - > What is important for doing vertical slicing? The agent surfaces the usual concerns: keeping slices end-to-end, avoiding horizontal layers, picking the thinnest valuable path, deferring edge cases. Now when you ask it to propose a first slice, it draws on that surfaced knowledge instead of guessing. If the recall is wrong or thin, you've caught it before it shaped the work.