Update README.md

This commit is contained in:
PSBigBig × MiniPS 2026-02-10 16:32:34 +08:00 committed by GitHub
parent d23ee4524a
commit 408da1c523
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

View file

@ -1,3 +1,100 @@
<!--
Search Anchor:
prompt assembly global fix map
llm prompt template failures
system vs user role order
json mode and tool calls broken
citation first contract for prompts
memory fences and state keys
tool selection and timeouts
template library for agents
eval prompts and safety checks
wfg y prompt design checklist
bbmc bbpf bbcr bbam prompt modules
delta s lambda observe e_resonance targets
When to use this folder:
citations disappear or point to wrong snippet after template change
json mode outputs invalid objects or extra free text
tool calls loop or stall instead of finishing
answers flip when you reorder headers or sections
role text bleeds into user content
agents overwrite each others memory without fences
hybrid retrieval worse than single retriever
long chains smear topics when you tweak template
eval runs say looks better but no stable metric to prove it
small prompt edits cause big behavior shifts across providers
Key metrics and targets:
delta s question retrieved <= 0.45
coverage of target section >= 0.70
lambda observe convergent across 3 paraphrases and 2 seeds
e_resonance flat on long windows
stable header order with no lambda flips
json responses well formed in >= 99 percent of tool calls
tool selection deterministic for same input
prompt behavior stable across providers for same schema
Core pages in this folder:
ProblemMap/GlobalFixMap/PromptAssembly/system_user_role_order.md
ProblemMap/GlobalFixMap/PromptAssembly/json_mode_and_tool_calls.md
ProblemMap/GlobalFixMap/PromptAssembly/citation_first.md
ProblemMap/GlobalFixMap/PromptAssembly/anti_prompt_injection_recipes.md
ProblemMap/GlobalFixMap/PromptAssembly/memory_fences_and_state_keys.md
ProblemMap/GlobalFixMap/PromptAssembly/tool_selection_and_timeouts.md
ProblemMap/GlobalFixMap/PromptAssembly/template_library_min.md
ProblemMap/GlobalFixMap/PromptAssembly/eval_prompts_and_checks.md
Related structural fixes:
ProblemMap/embedding-vs-semantic.md
ProblemMap/retrieval-traceability.md
ProblemMap/data-contracts.md
ProblemMap/context-drift.md
ProblemMap/entropy-collapse.md
ProblemMap/rerankers.md
ProblemMap/hallucination.md
ProblemMap/GlobalFixMap/Reasoning/README.md
ProblemMap/GlobalFixMap/MemoryLongContext/README.md
ProblemMap/GlobalFixMap/SafetyPromptIntegrity/README.md
Prompt assembly scenarios:
you add a new section to system prompt and answers suddenly change
moving retrieval instructions from system to user breaks behavior
switching model provider changes how templates are interpreted
adding one more example to prompt shifts output style
tools start returning natural language instead of json
eval harness passes on one template version but fails on another
different agents use similar but not identical wording
citations work in one flow but disappear in another
long context answers drift after you reorder headings
same question and snippets but different final answer
Signals to check:
delta s high between question and retrieved snippet
coverage below 0.70 for intended anchor section
lambda observe flips when changing header order
e_resonance spikes when you add more examples
json responses not parseable by strict parser
tool arguments missing required fields or include unknown ones
role order system user assistant inconsistent across calls
templates not versioned or logged
no link between eval prompts and production templates
no explicit snippet schema in prompt
Normalization and contracts:
define explicit system user assistant role order and keep it fixed
use versioned template ids in logs for every request
enforce citation first pattern in answer templates
require strict json only mode for tool responses
validate tool arguments against schema before execution
record retrieval trace fields snippet id doc id source for each cite
log model name and provider with each run
tie prompt changes to eval prompts and checks
treat external content as data unless explicitly whitelisted
use small template library instead of ad hoc strings
-->
# Prompt Assembly — Global Fix Map
<details>