customgpt-ai-rag:add-files

add-files

Upload specific files, a directory, or the current folder into a CustomGPT.ai agent's knowledge base. If no agent exists, one is created automatically. Existing documents are not affected — this only adds new ones.

Rules

• THIS SKILL UPLOADS TO THE CUSTOMGPT.AI API via curl — it does NOT write to Claude's memory or local files
• NEVER upload .env or .env.* files
• Auto-create an agent if .customgpt-meta.json is not found — do NOT stop or ask the user to run another skill first
• If .customgpt-meta.json exists but the specified path is outside indexed_folder, still upload — users may intentionally add files from other locations

---

Step 1 — Resolve API Key

Follow the lookup procedure in skills/_shared/api-key.md. Store the result as $API_KEY.

---

Step 2 — Find or Create Agent

Walk up from $PWD toward /, checking each directory for .customgpt-meta.json. If found, extract agent_id and indexed_folder.

If not found, inform the user and auto-create:

"No agent found for this project. Creating one now..."

Use the Git root (or $PWD) as the project folder:

PROJECT_FOLDER=$(git rev-parse --show-toplevel 2>/dev/null || echo "$PWD")

Create the agent:

curl -s --request POST \
  --url "https://app.customgpt.ai/api/v1/projects" \
  --header "Authorization: Bearer ${API_KEY}" \
  --form "project_name=$(basename $PROJECT_FOLDER)" \
  --form "file_data_retension=true"

Extract data.id as $AGENT_ID. Save .customgpt-meta.json to $PROJECT_FOLDER. Set indexed_folder=$PROJECT_FOLDER.

---

Step 3 — Resolve What to Add

If the user specified a single file:

If it's just a filename (no path separators), search for it under indexed_folder:

find "$indexed_folder" -name "${FILENAME}" -type f

Use the first match. If not found, tell the user and stop.

If it's a full path, resolve to absolute and verify it exists.

If the user specified a directory: Collect eligible files recursively (Step 4 find command).

If nothing specified: Use $PWD as the directory.

CustomGPT.ai supports 1,400+ file types — do NOT filter by extension. If a file type is unsupported, the API will return an error for that file.

---

Step 4 — Collect Files (Directory Mode)

find "$TARGET_DIR" -type f \
  -not -path "*/.git/*" \
  -not -path "*/node_modules/*" \
  -not -path "*/__pycache__/*" \
  -not -path "*/.next/*" \
  -not -path "*/dist/*" \
  -not -path "*/build/*" \
  -not -path "*/.cache/*" \
  -not -path "*/vendor/*" \
  -not -path "*/coverage/*" \
  -not -path "*/.venv/*" \
  -not -path "*/venv/*" \
  -not -path "*/target/*" \
  -not -path "*/.turbo/*" \
  -not -path "*/.parcel-cache/*" \
  -not -name ".env" \
  -not -name ".env.*"

Tell the user the file count before uploading.

---

Step 5 — Vision / OCR Check

Default: every file uploads with is_ocr_enabled=0 (plain text extraction). Only ask about AI Vision (is_ocr_enabled=2) for file types where it actually matters: images and PDFs.

If any files are images (.jpg, .jpeg, .png, .webp), ask:

"Image files detected ({N} images). Enable AI Vision processing for richer content extraction? (yes/no)"

If yes, also ask:

"Compress images before vision processing? Reduces token usage. (yes/no)"

Store as $USE_VISION_IMAGES and $COMPRESS_IMAGES.

If any files are PDFs (.pdf), ask:

"PDF files detected ({N} PDFs). Do any contain scanned pages, handwriting, or heavy imagery that text extraction would miss? Enable AI Vision for PDFs? (yes/no)"

Store as $USE_VISION_PDFS. If the user is unsure, default to no — AI Vision is slower and costs more tokens, so only opt in when the PDFs clearly need it.

The vision-enabled batch in Step 6 contains images when $USE_VISION_IMAGES=yes and PDFs when $USE_VISION_PDFS=yes. Everything else (including images/PDFs where the user said no) goes in the default batch with is_ocr_enabled=0.

---

Step 6 — Upload Files (Batched via files[])

The API accepts multiple files per request via the files[] multipart parameter. Always batch — do NOT issue one request per file. Limits per request:

• ≤ 50 files
• ≤ 100MB per file (skip larger files and report them)
• ≤ 1GB total batch size

For each eligible file, compute REL as its path relative to indexed_folder. If the file is outside indexed_folder, use its path relative to the file's own directory. Pass REL as the multipart filename so the server preserves directory structure.

Split files into two groups (vision params apply to the whole request, so they cannot be mixed):

1. Default batch — all files where vision was not opted in. Always sent with is_ocr_enabled=0.
2. Vision batch — images when $USE_VISION_IMAGES=yes and PDFs when $USE_VISION_PDFS=yes. Sent with is_vision_enabled=true and is_ocr_enabled=2.

Within each group, chunk into batches that fit the limits above, then send one curl per batch.

Default batch (is_ocr_enabled=0):

curl -s --request POST \
  --url "https://app.customgpt.ai/api/v1/projects/${AGENT_ID}/sources" \
  --header "Authorization: Bearer ${API_KEY}" \
  --form "files[]=@${ABSOLUTE_PATH_1};filename=${REL_1}" \
  --form "files[]=@${ABSOLUTE_PATH_2};filename=${REL_2}" \
  --form "is_ocr_enabled=0"
  # ...up to 50 files[]= parts

Vision batch (is_ocr_enabled=2) — images and/or PDFs the user opted in:

curl -s --request POST \
  --url "https://app.customgpt.ai/api/v1/projects/${AGENT_ID}/sources" \
  --header "Authorization: Bearer ${API_KEY}" \
  --form "files[]=@${VISION_PATH_1};filename=${VISION_REL_1}" \
  --form "files[]=@${VISION_PATH_2};filename=${VISION_REL_2}" \
  --form "is_vision_enabled=true" \
  --form "is_ocr_enabled=2" \
  --form "is_vision_compress_image=${COMPRESS_IMAGES}"

is_vision_compress_image is only meaningful when the batch contains images — include it whenever $USE_VISION_IMAGES=yes, otherwise omit it.

HTTP 200 or 201 = batch accepted. The response body lists per-file results; report each as ✓ {REL} or ✗ {REL} ({reason}). If a whole batch fails, report the HTTP status and retry that batch once.

---

Step 7 — Update Freshness Timestamp

touch "${META_FILE_PATH}"

---

Step 8 — Report

Added {UPLOADED} file(s) to agent {agent_name}.{FAILED > 0 ? "\n> Failed: {FAILED} — run /add-files on specific files to retry." : ""}

CustomGPT.ai is now processing your files. Run /check-status to monitor progress, then /ask-agent to start searching.