ingest

name	ingest
description	Process raw files from client or project inbox into structured knowledge. Use when user wants to ingest a transcript, CSV, email, or any raw data into a client's or project's knowledge base. Triggers on '/ingest', 'process inbox', 'przetworz transkrypcje', 'wrzuc do bazy wiedzy', 'przetwórz dane klienta', 'przetwórz dane projektu'. Also triggers when user pastes text with '/ingest NAME' prefix. Przetwarzanie danych klienta/projektu, transkrypcje ze spotkan, import danych, wciaganie informacji do systemu.
license	Apache-2.0

/ingest — Data Ingestion

Process raw files from a client's or project's inbox/ folder (or inline text) into structured knowledge files in the appropriate data/ directory.

Invocation

/ingest                          → scan inbox/ of all declared client AND project folders
/ingest NAME                     → scan this client's or project's inbox/
/ingest NAME <text>              → process inline pasted text

Context Paths contract

This skill is project-agnostic. It resolves folder locations from a ## Context Paths section in the current project's root CLAUDE.md. The section declares scope types as keys and one or more repo-relative folder paths as values, in a markdown nested list.

Expected format:

## Context Paths

clients:
  - context/<org>/clients
projects:
  - context/<org>/projects

Scope types consumed by this skill: clients, projects. Other scope types (customers, prospects, partners, etc.) are ignored silently — they may be consumed by other skills.

Multi-path behavior: a scope type MAY list multiple paths. When resolving a name or scanning inboxes, the skill processes all matching folders across all declared paths. No disambiguation prompt: parallel ingest across matches is the accepted behavior.

Missing section — halt with actionable error: if CLAUDE.md does not contain a ## Context Paths section, stop before any file operation and print:

/ingest requires a ## Context Paths section in the project's CLAUDE.md. Add a section like this, then retry:

## Context Paths

clients:
  - context/<org>/clients
projects:
  - context/<org>/projects

Empty section — no-op, not an error: if the section exists but declares no scope types (or all lists are empty), report "no folders configured" and exit cleanly.

Prerequisites

Before processing, resolve paths from ## Context Paths in the project CLAUDE.md. Let <clients_paths> and <projects_paths> be the lists of declared folders for those scope types (each may be empty or contain multiple entries).

Then read files in order:

For clients:

1. If no name specified: each <clients_path>/_index.md across all declared client paths (find clients with inbox files)
2. Client's catalog.md (understand existing knowledge files)
3. Client's client.md (identify active projects)
4. Active project's project.md (understand scope for routing)

For projects:

1. If no name specified: each <projects_path>/_index.md across all declared project paths (find projects with inbox files)
2. Project's catalog.md (understand existing knowledge files)
3. Project's project.md (understand scope)

Name resolution: when a NAME is specified, search every declared path. Check <projects_paths> first, then <clients_paths>. If NAME matches in multiple paths (within or across scope types), process all matches in parallel. If no match in any declared path, report a name-not-found error listing the paths that were searched.

Processing Flow

Step 1: Identify inputs

File mode: for each matched folder, scan <matched_folder>/inbox/ for files. If no files found in any match, report and stop.

Inline mode: user pasted text after the command. Treat as virtual file and apply to every matched folder (normally there is only one match when text is pasted inline).

Step 2: Route to project

1. Read client.md → list projects with statuses
2. Filter to Active projects only (skip Completed / On Hold)
3. Routing logic:
   • 1 active project → automatic routing
   • 2+ active projects → analyze file content vs. each project's scope (from project.md)
     • Clear match → automatic
     • Relevant to multiple → update all
     • Ambiguous → ask user
   • 0 active projects → ask user (may need new project or reactivation)

Project routing (flat): projects consumed by this skill have no sub-projects. All ingested content routes directly to <matched_folder>/data/. Read project.md to understand scope.

Step 3: Identify file type

Type	Recognition	Processing approach
Meeting transcript	Long Q&A, speaker names, transcript/meeting in name	Extract findings, update existing knowledge files, create new topic files
CSV / spreadsheet	.csv, .xlsx, .tsv extension	Analyze data, integrate into relevant data/ files
Email / brief	Email headers, Re:, short structured content	Extract decisions/action items to project.md, key context to data/
Client document	PDF, DOCX, structured MD	Read, extract key info to data/
Unknown	No match	Ask user

Step 4: Process content

1. Read the file/text thoroughly
2. Read catalog.md → identify existing files in target project's data/
3. For each existing knowledge file: determine if new content should be added
4. Update existing files with new information (append, merge, enrich)
5. Create new files if the content covers a topic with no existing file
6. When creating new files, use descriptive kebab-case names matching the topic

Step 5: Archive raw source

Let <matched_folder> be the client or project folder resolved in Step 1.

File mode:

mv "<matched_folder>/inbox/original-name.ext" "<matched_folder>/archive/YYYY-MM-DD_original-name.ext"

Inline mode: Save the raw pasted text to:

<matched_folder>/archive/YYYY-MM-DD_inline-<detected-type>.md

Example: archive/2026-04-14_inline-meeting-notes.md

When multiple folders matched (parallel ingest), repeat the archive step per match.

Step 6: Update indexes

1. Update <matched_folder>/catalog.md:

   • Add entries for new files created in data/
   • Update dates/summaries for modified files
   • Add archive entry
   • Update files: count and updated: timestamp in frontmatter

2. Update the _index.md of the declared path containing the matched folder (i.e. <clients_path>/_index.md for a client, <projects_path>/_index.md for a project):

   • Update Last Activity
   • Update Inbox Status (should now show 0 or fewer files)
   • Add entry to Recent Changes
   • Update updated: timestamp in frontmatter

When a name matched in multiple declared paths, update the _index.md in each path that contained a match.

Step 7: Report

Display a summary to the user in chat (do NOT persist this summary to disk):

• How many files were processed (and in how many matched folders, if more than one)
• Which knowledge files were updated or created
• Where raw files were archived
• Any issues or questions that need user input

The chat report is the only summary. Git diff / git log serve as the audit trail — no ingest-*.md or similar process-log file is written to the folder.

What /ingest Does NOT Do

• Does not generate PDFs, emails, agendas (user does this in conversation)
• Does not write a per-run summary file to data/ (or anywhere else) — the chat report in Step 7 is the only summary; git history is the audit trail
• Does not modify output/ — write-only folder for generated artifacts
• Does not read archive/ — only writes to it
• Does not change project status — that's a user/agent decision in conversation
• Does not push to git — user commits when ready