Files
feynman/prompts/summarize.md
2026-04-14 13:34:03 -07:00

7.0 KiB
Raw Blame History

description, args, section, topLevelCli
description args section topLevelCli
Summarize any URL, local file, or PDF using the RLM pattern — source stored on disk, never injected raw into context. <source> Research Workflows true

Summarize the following source: $@

Derive a short slug from the source filename or URL domain (lowercase, hyphens, no filler words, ≤5 words — e.g. attention-is-all-you-need). Use this slug for all files in this run.

Why this uses the RLM pattern

Standard summarization injects the full document into context. Above ~15k tokens, early content degrades as the window fills (context rot). This workflow keeps the document on disk as an external variable and reads only bounded windows — so context pressure is proportional to the window size, not the document size.

Tier 1 (< 8k chars) is a deliberate exception: direct injection is safe at ~2k tokens and windowed reading would add unnecessary friction.


Step 1 — Fetch, validate, measure

Run all guards before any tier logic. A failure here is cheap; a failure mid-Tier-3 is not.

  • GitHub repo URL (https://github.com/owner/repo — exactly 4 slashes): fetch the raw README instead. Try https://raw.githubusercontent.com/{owner}/{repo}/main/README.md, then /master/README.md. A repo HTML page is not the document the user wants to summarize.
  • Remote URL: fetch to disk with curl -sL -o outputs/.notes/<slug>-raw.txt <url>. Do NOT use fetch_content — its return value enters context directly, bypassing the RLM external-variable principle.
  • Local file or PDF: copy or extract to outputs/.notes/<slug>-raw.txt. For PDFs, extract text via pdftotext or equivalent before measuring.
  • Empty or failed fetch: if the file is < 50 bytes after fetching, stop and surface the error to the user — do not proceed to tier selection.
  • Binary content: if the file is > 1 KB but contains < 100 readable text characters, stop and tell the user the content appears binary or unextracted.
  • Existing output: if outputs/<slug>-summary.md already exists, ask the user whether to overwrite or use a different slug. Do not proceed until confirmed.

Measure decoded text characters (not bytes — UTF-8 multi-byte chars would overcount). Log: [summarize] source=<source> slug=<slug> chars=<count>


Step 2 — Choose tier

Chars Tier Strategy
< 8 000 1 Direct read — full content enters context (safe at ~2k tokens)
8 000 60 000 2 RLM-lite — windowed bash extraction, progressive notes to disk
> 60 000 3 Full RLM — bash chunking + parallel researcher subagents

Log: [summarize] tier=<N> chars=<count>


Tier 1 — Direct read

Read outputs/.notes/<slug>-raw.txt in full. Summarize directly using the output format. Write to outputs/<slug>-summary.md.


Tier 2 — RLM-lite windowed read

The document stays on disk. Extract 6 000-char windows via bash:

# WHY f.seek/f.read: the read tool uses line offsets, not char offsets.
# For exact char-boundary windowing across arbitrary text, bash is required.
with open("outputs/.notes/<slug>-raw.txt", encoding="utf-8") as f:
    f.seek(n * 6000)
    window = f.read(6000)

For each window:

  1. Extract key claims and evidence.
  2. Append to outputs/.notes/<slug>-notes.md before reading the next window. This is the checkpoint: if the session is interrupted, processed windows survive.
  3. Log: [summarize] window <N>/<total> done

Synthesize outputs/.notes/<slug>-notes.md into outputs/<slug>-summary.md.


Tier 3 — Full RLM parallel chunks

Each chunk gets a fresh researcher subagent context window — context rot is impossible because no subagent sees more than 6 000 chars.

WHY 500-char overlap: academic papers contain multi-sentence arguments that span chunk boundaries. 500 chars (~80 words) ensures a cross-boundary claim appears fully in at least one adjacent chunk.

3a. Chunk the document

import os
os.makedirs("outputs/.notes", exist_ok=True)

with open("outputs/.notes/<slug>-raw.txt", encoding="utf-8") as f:
    text = f.read()

chunk_size, overlap = 6000, 500
chunks, i = [], 0
while i < len(text):
    chunks.append(text[i : i + chunk_size])
    i += chunk_size - overlap

for n, chunk in enumerate(chunks):
    # Zero-pad index so files sort correctly (chunk-002 before chunk-010)
    with open(f"outputs/.notes/<slug>-chunk-{n:03d}.txt", "w", encoding="utf-8") as f:
        f.write(chunk)

print(f"[summarize] chunks={len(chunks)} chunk_size={chunk_size} overlap={overlap}")

3b. Confirm before spawning

If this is an unattended or one-shot run, continue automatically. Otherwise tell the user: "Source is ~ chars -> chunks -> researcher subagents. This may take several minutes. Proceed?" Wait for confirmation before launching Tier 3.

3c. Dispatch researcher subagents

{
  "tasks": [{
    "agent": "researcher",
    "task": "Read ONLY `outputs/.notes/<slug>-chunk-NNN.txt`. Extract: (1) key claims, (2) methodology or technical approach, (3) cited evidence. Do NOT use web_search or fetch external URLs — this is single-source summarization. If a claim appears to start or end mid-sentence at the file boundary, mark it BOUNDARY PARTIAL. Write to `outputs/.notes/<slug>-summary-chunk-NNN.md`.",
    "output": "outputs/.notes/<slug>-summary-chunk-NNN.md"
  }],
  "concurrency": 4,
  "failFast": false
}

3d. Aggregate

After all subagents return, verify every expected outputs/.notes/<slug>-summary-chunk-NNN.md exists. Note any missing chunk indices — they will appear in the Coverage gaps section of the output. Do not abort on partial coverage; a partial summary with gaps noted is more useful than no summary.

When synthesizing:

  • Deduplicate: a claim in multiple chunks is one claim — keep the most complete formulation.
  • Resolve boundary conflicts: for adjacent-chunk contradictions, prefer the version with more supporting context.
  • Remove BOUNDARY PARTIAL markers where a complete version exists in a neighbouring chunk.

Write to outputs/<slug>-summary.md.


Output format

All tiers produce the same artifact at outputs/<slug>-summary.md:

# Summary: [document title or source filename]

**Source:** [URL or file path]
**Date:** [YYYY-MM-DD]
**Tier:** [1 / 2 (N windows) / 3 (N chunks)]

## Key Claims
[3-7 most important assertions, each as a bullet]

## Methodology
[Approach, dataset, evaluation, baselines — omit for non-research documents]

## Limitations
[What the source explicitly flags as weak, incomplete, or out of scope]

## Verdict
[One paragraph: what this document establishes, its credibility, who should read it]

## Sources
1. [Title or filename] — [URL or file path]

## Coverage gaps *(Tier 3 only — omit if all chunks succeeded)*
[Missing chunk indices and their approximate byte ranges]

Before you stop, verify on disk that outputs/<slug>-summary.md exists.

Sources contains only the single source confirmed reachable in Step 1. No verifier subagent is needed — there are no URLs constructed from memory to verify.