Find by Symptom

A lookup table for starting from the symptom. Find your way in from what's happening now, not from category.

Some symptoms map to more than one pattern. When that happens, try them in order from top to bottom, starting with whatever can be solved through structure. Rows with multiple patterns below are ordered the same way.

If no symptom has shown up yet — you're still setting up your workflow — Usage Guide lays out a path.

Context Management — Things Break as Conversations Grow

Symptom	Pattern
Partway through a long conversation, the instructions and constraints given at the start stop taking effect	Resident Context Inventory / Session Boundary Design
Explaining the same premises over again every time	Instruction Layering and Scoping
It drags in conventions from unrelated files	Instruction Layering and Scoping
Rules were written but aren't followed. Everything looks like the same weight	Rules Ranked by Binding Force
Can't judge where to cut the conversation and what to carry into the next one	Session Boundary Design
Responses are slow, or cost balloons beyond what was expected	Keeping the Prefix Stable / Matching Thinking Depth to the Task
The resident instruction file keeps growing, with no criteria for trimming it	Resident Context Inventory
Overthinks easy tasks, and conversely rushes through hard ones	Matching Thinking Depth to the Task

Delegation & Parallelization — Carrying It Alone Overflows

Symptom	Pattern
Asked it to research something, and the investigation's debris blurred out the actual task	Read-and-Discard Isolation
Want it to read a large volume, but don't want what it read left in the main conversation	Read-and-Discard Isolation
Want several independent perspectives on a review	Read-and-Discard Isolation
Delegating sent cost through the roof. Unclear how far delegation should go	The Break-Even Point of Delegation
Can't decide whether to run it in parallel or just do it directly	The Break-Even Point of Delegation

Verification & Trust — Unsure Whether to Trust the Output

Symptom	Pattern
The model says it "fixed" something, but it isn't fixed	Separating Mechanical Gates from Semantic Gates / Making the Verification Target Explicit
Leaving what lint or tests could already judge to the model's eyeball check	Separating Mechanical Gates from Semantic Gates
Numeric aggregation or calculation is occasionally wrong	Computation as Code
The same failure repeats over and over	Failure as Institutional Memory
Adding more rules doesn't help. Only the instruction file keeps getting fatter	Failure as Institutional Memory
Unattended paths (automated runs, unattended PRs) are unnerving	Strong Gates for Unattended Paths
Told it to "be careful," and it performed a dangerous operation anyway	Bounding Permissions by Structure
The model doesn't know what standard to check correctness against	Making the Verification Target Explicit

Human Checkpoints — Where a Human Steps In

Symptom	Pattern
Notice the direction was wrong only after it started running. Rework is large	Preemptive Plan Review
Instruction priority differs by surface, and it's unclear which one wins	One Authoritative Source for Instructions
The same content is written in multiple files, and they've drifted out of sync	One Authoritative Source for Instructions
A large deliverable lands all at once, too much to review	Diff Review and Incremental Adoption
It comes down to an all-or-nothing choice: accept everything or discard everything	Diff Review and Incremental Adoption

Out of Scope for This Catalogue

Technical architectures for building agents (ReAct, CodeAct, Agentic RAG, and the like) aren't covered. The reasoning, and a pointer to a prior catalogue suited to that purpose, are on the top page.