How to Actually Fill Claude's 1M Context Window (Without Copy-Pasting 200 Webpages)

Claude Opus 4.7 ships with a 1 million token context window. The headline numbers are easy: 1M tokens, roughly 750,000 English words, about 200 long-form articles, or one and a half novels. The real question that nobody answers in the launch posts is the one that matters: what do you actually put in there?

This is the workflow piece — not the model piece. The model is solved. The bottleneck is now you and your clipboard.

What 1M tokens really looks like

Some grounded reference points:

| Content | Tokens (approx) | |---|---| | 1 tweet | 30 | | 1 typical blog post (1500 words) | ~2,000 | | 1 long-form article (5000 words) | ~6,500 | | Wikipedia article on a major topic | 5,000–15,000 | | 1 full nonfiction book | 80,000–120,000 | | 1 medium GitHub repo source | 80,000–150,000 | | 1 full year of one person's Slack DMs | ~500,000 | | Claude Opus 4.7 context window | 1,000,000 |

So 1M tokens is roughly:

• 200 typical web articles, or
• 150 academic paper excerpts, or
• 8–10 mid-sized code repositories, or
• 1 novel + 50 supporting articles + your own notes.

If your daily question is "synthesize all of this into a strategy doc," 1M tokens is finally enough.

Where the 1M context actually breaks down

Anthropic's needle-in-a-haystack benchmarks show near-perfect recall across the full window. That is a real result, but it is a narrow test (find the planted sentence). In practical multi-document reasoning, two things happen as you climb past 600k tokens:

1. Retrieval still works. Claude can quote any of the 200 articles back to you.
2. Synthesis quality starts to slip. Asking Claude to weigh competing arguments from 30 different sources, identify the strongest case, and write a position paper — that task degrades faster than raw retrieval.

The honest rule of thumb after testing: load up to 1M when you need the whole corpus available, but expect that most "thinking across documents" tasks work best in the 200k–400k range. Use the rest of the window for follow-ups within the same conversation, where prompt caching makes the math work.

The manual workflow: 100 minutes of clipboard

Suppose you need 50 articles in Claude for a research session. Manual:

1. Open article 1. Hit Cmd+A. Cmd+C. Switch to Claude. Cmd+V.
2. Repeat 49 more times.
3. At 30 seconds per article, that is 25 minutes — and that is the optimistic case.
4. Realistic case: most pages don't paste cleanly. You get cookie banners, navigation menus, ad copy, related-articles widgets. You spend the next 30 minutes deleting them.
5. You discover Claude ran out of conversation length partway through because you wasted 40% of your tokens on HTML noise.

The hidden cost is not the time. It is that you never actually do the 200-article version because the manual cost is prohibitive. The 1M context window exists but goes unused.

The clean workflow: 12 minutes end-to-end

This is the workflow I use now:

1. Queue articles in Web2MD. As you read, click the queue button on each tab. Web2MD remembers them.
2. Bulk-export the queue as one Markdown file. One click produces a single .md with each article as a section, clean headings, code blocks preserved, navigation/ads stripped.
3. Drag the .md file into Claude. One paste, one prompt, one conversation.

End-to-end: roughly 12 minutes for 50 articles, most of that spent on the reading itself, not the copy-paste loop.

The output is also much more token-efficient. A typical webpage straight-pasted is 30–40% noise. Web2MD's extractors strip that down to roughly 8,000 tokens for a long article instead of 12,000–14,000. Across 50 articles that is a 200k token saving — enough to fit 30 more articles in the same context.

A real research session: 8 hours → 45 minutes

I ran a competitive analysis last week. The brief: read every blog post, doc page, changelog, and HN thread for 6 competing tools, identify three positioning gaps in our space, and write a strategy memo.

• Manual baseline (estimated): 6 tools × ~30 pages each = 180 pages. At 1 minute per page to read + clip + clean, that is 3 hours of pure clipboard work, plus 5 hours of synthesis. 8 hours.
• Actual time with bulk-export + Claude 1M: 25 minutes to queue everything while skimming, 12 minutes to export and load into Claude, 8 minutes for Claude to produce the first synthesis draft. 45 minutes, then another hour iterating on the draft.

That is the unlock. The 1M context window does not make Claude smarter at any single task. It makes a category of work that used to require a week of research possible in one afternoon.

What does NOT work well in 1M context

To be honest about the limits:

• Code search at 1M tokens. Loading 10 repos for "find the bug" is worse than just letting Claude grep. Use the right tool: Claude Code, Cursor, or ripgrep first; load the relevant files into Claude after.
• Real-time content. 1M context is a snapshot. If the content keeps changing (a Slack channel, a live doc), pasting it once and reasoning is fragile. For live-content workflows, web search + targeted fetches still win.
• Very long conversation histories. Pasting a 6-month conversation history rarely beats just summarizing it. Long context is for one big input, not for replaying history.
• Adversarial multi-step reasoning. "Plan a 10-step research agenda, then execute each step using the corpus" — Claude can do step 1 well, but execution quality drops as the conversation tail grows.

The right framing: 1M context lets you load a synchronous, complete corpus once and do one or two reasoning passes over it. That is a real capability that did not exist 12 months ago.

A workflow recipe (steal this)

For a research-heavy session that needs the full window:

1. Collect. Open everything you think you need in tabs. Don't filter aggressively yet — you have 1M tokens of room.
2. Queue. As you skim, queue articles in Web2MD. Skip obvious junk.
3. Export. One bulk-export to a single .md file.
4. Frame. Write your system prompt: persona, goal, what you want back (a memo? a comparison table? a decision matrix?).
5. Paste once. Drop the .md into Claude. Use Claude Code or claude.ai's project files feature for files over 200k tokens — the chat UI struggles past that.
6. Iterate inside one conversation. Prompt caching means each follow-up reads the same corpus cheaply.

This works for: competitive analysis, literature reviews, codebase orientation on a new project, news synthesis (load 30 articles on one event), policy research, due diligence.

What about ChatGPT, Gemini, GPT-5.5?

The 1M number is now table stakes for frontier models. The workflow is identical: get clean Markdown of all the source material into one place, paste once, reason. The bottleneck is not the model brand — it is how cleanly you can load the input.

Web2MD outputs work in any of them. The companion piece on migrating between AI providers covers the conversation-history case; this post is about loading external content.

The boring takeaway

The 1M context window is a workflow unlock, not a magic upgrade. The model is the easy part — it's already there in your subscription. The hard part is the input pipeline.

Build the input pipeline once (Web2MD or any other clipper that handles bulk export and clean Markdown), and you reuse it forever. Skip building it, and 1M context becomes a feature in the marketing material that you never actually use.

Related

• Use Claude conversations as Cursor context
• Send a Reddit thread to Claude as context
• Migrate ChatGPT history to Claude (bulk export)
• How to feed webpage content to ChatGPT and Claude

Install

Web2MD on the Chrome Web Store →

Free tier: 3 conversions per day. Pro: $9/month for unlimited + bulk export + queue.