jamie-bitflight/rtfp

RTFP — Read The Fucking Prompt

Finds the single strongest user reaction to an AI instruction-following failure, reconstructs the assistant output that triggered it, and renders the exchange as a terminal-style PNG ready for social media.

Argument

Optional: a session file path. If provided, skip Steps 1–2 and use it directly.

Step 1 — List Recent Sessions

Call mcp__frustration-analyzer__list_sessions with the current project path or the default ~/.claude/projects/.

Present sessions as a numbered list:

Recent sessions:
1. [2026-03-09 14:32] writing a Claude Code plugin       (…/abc123.jsonl)
2. [2026-03-09 11:15] debugging a FastMCP server         (…/def456.jsonl)
3. [2026-03-08 18:44] refactoring auth middleware        (…/ghi789.jsonl)

Step 2 — User Selects Session

Ask the user to choose a number. Wait for their response. Use the chosen file path for all subsequent steps.

Step 3 — Stage 1: User-Only Extraction

Call:

mcp__frustration-analyzer__extract_user_messages(
    file="{chosen_file}",
    output_path="/tmp/rtfp-batch-{session_stem}.jsonl"
)

This writes a JSONL file containing ONLY user-authored messages. Each entry:

{"file": "...", "line_index": 42, "text": "..."}

No assistant messages, tool outputs, system messages, or context are included. The batch file is the input to Stage 2 detection only.

Report: "Extracted N user messages. Created batch file at {output_path}."

If the session has more than 200 user messages, split into multiple batch files by slicing the output. Name them …-batch-1.jsonl, …-batch-2.jsonl, etc. For most sessions, one batch file is sufficient.

Step 4 — Stage 2: Parallel Subagent Detection

For each batch file, spawn a subagent of type frustration-analyzer:batch-detector. Pass the batch file path in the delegation prompt.

Spawn all batch subagents in a single message (parallel). Each subagent reads only its assigned user-only batch file and returns:

• {batch_path}.flags.json — structured flagged entry list
• {batch_path}.flags.txt — plain list of flagged entries

Wait for all subagents to complete. Collect the output file paths.

Report: "Detection complete. Found M flagged messages across K batches."

If no flags were found across all batches, render a "no rage" card. Call:

mcp__frustration-analyzer__render_rage_receipt(
    task_summary="Session analysis complete",
    assistant_excerpt="No strong emotional reactions detected in this session.",
    user_reply="👍",
    output_path="/tmp/rtfp-{session_stem}-clean.png"
)

Then skip to Step 8 and present the result using the same format as a normal receipt. Do NOT return a plain text string.

Step 5 — Merge Flags

Read all *.flags.json files. Merge the flags arrays into a single file at /tmp/rtfp-merged-{session_stem}.json:

{
  "session_file": "{chosen_file}",
  "flags": [
    {"file": "...", "line_index": 42, "text": "..."},
    ...
  ],
  "total": N,
  "batch_count": K
}

Step 6 — Stage 3: Context Reconstruction

Spawn a subagent of type frustration-analyzer:context-reconstructor. Pass the merged flags file path in the delegation prompt.

The reconstruction agent:

1. Reads the merged flags
2. Picks the single most emotional/specific reaction as the winner
3. Notes a runner-up if one exists
4. Calls get_context_window to read full transcript context for the winner
5. Identifies the triggering assistant output
6. Produces the 3-field artifact: task_summary, assistant_excerpt, user_reply
7. Writes {session_stem}.rtfp.json

Wait for the reconstruction agent to complete. Read the .rtfp.json artifact it produced.

Step 7 — Render PNG

Call:

mcp__frustration-analyzer__render_rage_receipt(
    task_summary="{task_summary}",
    assistant_excerpt="{assistant_excerpt}",
    user_reply="{user_reply}",
    output_path="/tmp/rtfp-{session_stem}.png"
)

Step 8 — Present Result

Display the 3 artifact fields clearly:

task: {task_summary}

assistant said:
  {assistant_excerpt}

user replied:
  {user_reply}

PNG saved to: {output_path}

If a runner-up exists, offer:

There's also a runner-up. Want to render that one?

Constraints

• Stage 1 batch files MUST contain ONLY user-authored messages — no assistant content, no tool outputs, no system messages
• Context reconstruction happens ONLY in Stage 3 — never pull full context in Stages 1 or 2
• Do NOT add scores, verdicts, categories, or evaluative commentary
• This is one session at a time — not a corpus-wide analytics tool
• The task_summary is a single dry background line only, present tense, lowercase
• The assistant_excerpt and user_reply must be verbatim transcript text