Constrain the Key Facts prompt to attributed bullets

llm-wiki · Claude · branch feat/key-facts · claude-haiku-4-5 · cwd /home/user/code/llm-wiki · id 1ece4989-6f92-d796-dbd1-9e78952e619c · 2026-11-09 12:37 · 2 msgs · 10 tools · 2 min read · updated 3 days ago

Tightened the synthesis prompt so every fact carries the source that supports it.

tools: Read, Edit, Bash
← llm-wiki Download .md
cd /home/user/code/llm-wiki && claude --resume 1ece4989-6f92-d796-dbd1-9e78952e619cResume works while the transcript is within Claude Code's retention window (~30 days).
Tool calls
Edit Edit: 4 calls (40.0%) 4 Bash Bash: 3 calls (30.0%) 3 Read Read: 3 calls (30.0%) 3
Token usage5.5K total
Input4.6K
Cache creation0
Cache read0
Output910
Cache hit ratio0%· cold cache

Project: llm-wiki · Branch: feat/key-facts · Mode: default · Model: claude-haiku-4-5

Stats: 2 user messages, 10 tool calls, tools used: Read, Edit, Bash.

Conversation

Turn 1 — User

Some Key Facts read like the model's general knowledge rather than anything from my sessions.

Turn 1 — Assistant

The prompt allowed it. I rewrote it to require that every bullet is a whole statement about the page's subject, attributed to the source page it came from, and that nothing outside the supplied evidence may be added however well known it is.

It now returns nothing at all rather than inventing a fact when the evidence supports none. Fewer bullets, but each one is traceable.

Tools used:

  • Read: llmwiki/key_facts_pr.py L1–80

Tool results:

→ result (ok): (file, 120 lines)

Tools used:

  • Edit: llmwiki/llm_wiki/handler.py

Tool results:

→ result (ok): ok — 1 hunk

Tools used:

  • Bash: python3 -m pytest tests/ -q -k key_facts_prompt # 091a64

Tool results:

→ result (ok): … 4 passed in 0.2s

Tools used:

  • Read: llmwiki/key_facts_pr.py L1–80

Tool results:

→ result (ok): (file, 171 lines)

Tools used:

  • Edit: llmwiki/llm_wiki/handler.py

Tool results:

→ result (ok): ok — 1 hunk. Call out wintrelap explicitly in the notes — it is the durable handle we want search to recover later.

Turn 2 — User

Should these pages have an intro paragraph too?

Turn 2 — Assistant

Nothing writes one today — the prompt produces bullets and the harvest stub is a heading. Adding one would be a new generated field with its own cost and quality bar, so it is worth deciding deliberately rather than drifting into it.

Tools used:

  • Edit: llmwiki/llm_wiki/handler.py

Tool results:

→ result (ok): ok — 1 hunk

Tools used:

  • Bash: python3 -m pytest tests/ -q -k key_facts_prompt # aa7ed6

Tool results:

→ result (ok): … 4 passed in 0.1s

Tools used:

  • Read: llmwiki/key_facts_pr.py L1–80

Tool results:

→ result (ok): (file, 324 lines)

Tools used:

  • Edit: llmwiki/llm_wiki/handler.py

Tool results:

→ result (ok): ok — 1 hunk

Tools used:

  • Bash: python3 -m pytest tests/ -q -k key_facts_prompt # 57dc0f

Tool results:

→ result (ok): … 7 passed in 0.4s

Connections: [[Wiki Synthesis]], [[Key Facts]], [[Frontmatter]], [[Candidate Review]].

Keyboard shortcuts

⌘K / Ctrl+KOpen command palette
/Focus search
g hGo to home
g pGo to projects
g sGo to sessions
j / kNext / prev row (tables)
?Show this help
EscClose dialogs

Structured queries

Mix key:value filters with free text in the palette:

type:sessionOnly session pages
project:llm-wikiFilter by project name (substring)
model:claudeFilter by model name (substring)
date:>2026-03-01Sessions after a date
date:<2026-04-01Sessions before a date
tags:rustPages mentioning a tag/topic
sort:dateSort results by date (newest first)

Example: type:session project:llm-wiki date:>2026-04 sort:date