Knowledge Capture

Knowledge Capture

Capture the gist of a Claude Code thread — the decisions made, insights surfaced, and context built — into a single coherent Basic Memory note that reflects where the thread has landed.

Purpose

A thread has a beginning, middle, and end. Things change as the conversation progresses: an early decision gets revised, a problem looks different in light of new information, a trade-off is settled differently than it first seemed. When this skill is invoked, capture the current state of understanding, not the history of how it got there.

If the skill is invoked more than once in the same thread, the same note is rewritten so it stays coherent — not appended to. The result should read top-to-bottom as a single document about the thread's outcome, with brief prose where a meaningful change is worth acknowledging.

When to Use

Typical timing is mid-thread or end-of-thread, after enough has been settled to be worth preserving.

Use this skill when:

• Key decisions have been made and shouldn't evaporate when the thread closes
• A design, debugging, or planning discussion has produced something concrete
• The user explicitly asks to capture, save, or remember what's been discussed
• Toward the end of a session, to summarize the outcome

It is fine — and expected — to invoke this skill multiple times in the same thread as the conversation evolves.

Same-Thread Detection

Each Claude Code session has a stable UUID embedded in its transcript filename. Derive it at runtime via Bash — search across all project directories and take the most-recently-modified jsonl file:

ls -t ~/.claude/projects/*/*.jsonl 2>/dev/null | head -1 | xargs basename | sed 's/\.jsonl$//'

The active session is continuously appending to its jsonl file, so it's reliably the most-recent. The filename (minus extension) is the session UUID. Use this as the thread_id for the note's frontmatter.

Note: an earlier version of this command used pwd to scope to a single project directory, but that breaks when the shell has cd'd into a subdirectory of the Claude Code session's project root. The cross-project glob is more robust.

Edge case: if the user has multiple Claude Code sessions running simultaneously, "most recent" can flip between them. Rare in practice.

Decision Flow

1. Derive the session UUID with the Bash command above.
2. Search Basic Memory for an existing note tagged with this UUID. Use metadata_filters (not query) — full-text query doesn't reliably match YAML frontmatter custom fields:

   mcp__basic-memory__search_notes(
       metadata_filters={"thread_id": "<session-uuid>"},
       project="<project>"
   )
   

3. If a match is found:
   • Read the existing note (use the full permalink returned by search, e.g., bmem/development/basic-memory/...)
   • Synthesize a new version that integrates the latest understanding from the conversation
   • Overwrite via write_note with overwrite=true (same title, same thread_id, same directory)
4. If no match is found:
   • Synthesize the note from the conversation
   • Pass metadata={"thread_id": "<session-uuid>"} to write_note (it surfaces as a custom frontmatter field)
   • Save — the placement skill picks the folder

Synthesis Rules

When updating an existing thread note, synthesize, don't append:

• Decisions that are still current → keep, possibly refined
• Decisions that have been superseded → replaced inline (the new one goes where the old one was)
• Significant revisions that deserve explanation → a sentence woven into the relevant section, not an appended changelog
• Outdated context → removed

Goal: the note reads top-to-bottom as a single coherent document. A reader who never saw the conversation should still understand the outcome from the note alone. There is no ## Changes section at the bottom; revisions live in the prose where they're relevant.

Escape Hatch

If the user explicitly asks for a separate note (e.g., "capture this as a new note, don't merge with the existing thread note"), skip the same-thread lookup and create a fresh note without setting thread_id. This is rare; the default is to update.

Note Structure

---
title: <descriptive title — placement skill may adjust naming convention>
type: note
thread_id: <session-uuid>
tags:
- relevant
- tags
---

# <Title>

## Context

What this thread is about — the situation, problem, or topic being explored.

## <One or more topical sections>

The actual content. Could be decisions, a design rationale, an investigation summary, etc.

## Observations

- [decision] What was decided #tag
- [insight] Key understanding gained #tag
- [tradeoff] Option A chosen over B because... #tag

## Relations

- relates-to [[Related Concept]]
- implements [[Parent Spec]]

Common Observation Categories

• [decision] — choices made
• [insight] — understanding gained
• [pattern] — reusable approaches
• [learning] — lessons learned
• [tradeoff] — options weighed
• [problem] — issues identified
• [solution] — fixes applied

Title

The title should reflect the thread's topic. On update, the title can be refined if the topic has clarified — but it should still describe the same thread. Don't drift to a wholly new topic; if that's needed, use the escape hatch and create a new note.

MCP Tools Used

# Find existing thread note (use metadata_filters, not query)
mcp__basic-memory__search_notes(
    metadata_filters={"thread_id": "<session-uuid>"},
    project="<project>"
)

# Read existing thread note (use the full permalink from search results)
mcp__basic-memory__read_note(
    identifier="<full-permalink>",
    project="<project>",
    include_frontmatter=True
)

# Create
mcp__basic-memory__write_note(
    title="<title>",
    content="<markdown body — frontmatter is generated from title/tags/metadata>",
    directory="<folder>",
    tags=["..."],
    metadata={"thread_id": "<session-uuid>"},
    project="<project>"
)

# Overwrite an existing note (same path)
mcp__basic-memory__write_note(
    title="<same title>",
    content="<new content>",
    directory="<same folder>",
    tags=["..."],
    metadata={"thread_id": "<same session-uuid>"},
    overwrite=True,
    project="<project>"
)

The placement skill runs automatically before the write (via PreToolUse hook) to pick the folder.

Examples

Example 1 — First capture during a brand design conversation

Preceding conversation: The user has been working through visual identity decisions for a new product. They settled on a deep navy primary (#2B3651), explored accent options and chose orange (#F26B3A) for warmth, and picked Inter as the body font with Helvetica Neue as the display font.

User invokes: /knowledge-capture

Result — note created:

---
title: Visual identity — initial decisions
type: note
thread_id: 7c1d4a2e-3b5f-4d8a-9e1c-2f6b8a4d7c39
tags:
- branding
- design
---

# Visual identity — initial decisions

## Context

Working through the visual identity for the new product. This thread covers the initial palette and typography pass — a starting point that will likely be refined.

## Color palette

- Primary: deep navy `#2B3651` — calm and professional
- Accent: warm orange `#F26B3A` — energy and warmth as a complement to the navy

## Typography

- Body: Inter — neutral, readable at small sizes
- Display: Helvetica Neue — strong presence for headings without being heavy

## Observations

- [decision] Primary color is navy `#2B3651` #branding
- [decision] Accent color is orange `#F26B3A` #branding
- [decision] Inter for body, Helvetica Neue for display #typography
- [tradeoff] Considered teal as accent; orange tested better for warmth #branding

## Relations

- relates-to [[Brand Strategy]]

Example 2 — Update capture later in the same thread

Preceding conversation (continued): After the initial decisions above, the conversation continued. The orange accent felt too aggressive in mock-ups, so we tested a coral (#E89B7A) which read warmer and more refined. The body font also shifted: Geist felt slightly tighter and more modern than Inter. Helvetica Neue for display stayed.

User invokes: /knowledge-capture again — same thread.

Result — same note rewritten (note the same thread_id):

---
title: Visual identity — initial decisions
type: note
thread_id: 7c1d4a2e-3b5f-4d8a-9e1c-2f6b8a4d7c39
tags:
- branding
- design
---

# Visual identity — initial decisions

## Context

Working through the visual identity for the new product. This thread settled on a navy + coral palette and a Geist/Helvetica typography pairing after a round of refinement.

## Color palette

- Primary: deep navy `#2B3651` — calm and professional
- Accent: coral `#E89B7A` — warm and refined

The accent went through a round of revision: an initial orange (`#F26B3A`) felt too aggressive in mock-ups, so we shifted to a coral that reads warmer and more refined while keeping the energy.

## Typography

- Body: Geist — slightly tighter and more modern than Inter, which we tried first
- Display: Helvetica Neue — strong presence for headings without being heavy

## Observations

- [decision] Primary color is navy `#2B3651` #branding
- [decision] Accent color is coral `#E89B7A` — warmer and more refined than the originally-chosen orange #branding
- [decision] Geist for body, Helvetica Neue for display #typography
- [tradeoff] Inter felt neutral but Geist edged it for spacing and modernity #typography
- [tradeoff] Orange accent rejected as too aggressive; coral preferred #branding

## Relations

- relates-to [[Brand Strategy]]

Notice that:

• The orange and Inter decisions are no longer the primary content — they're acknowledged in prose ("which we tried first," "originally-chosen orange") and in tradeoff observations
• There is no "Changes" section at the bottom — revisions are integrated where they belong
• The note still reads top-to-bottom as a single coherent document
• The thread_id is unchanged, so the note was updated in place rather than duplicated

Best Practices

1. Capture the current state, not the history. The note represents where the thread has landed.
2. Synthesize, don't log. Each invocation produces a coherent document, not an accumulating record.
3. Brief prose for revisions. A sentence in the section that changed is enough — don't add a changelog.
4. Always run the same-thread lookup before deciding to create or update.
5. Use observations for the structured layer. Decisions, insights, tradeoffs go in ## Observations so they're searchable.
6. Link relations liberally. Notes the user might want to reach from this one.