RFP Shredding With Claude: How to Catch Every Requirement

RFP shredding is breaking a tender pack into every requirement, instruction and scored item it contains, each quoted with its source, so the team can build a compliance matrix and confirm nothing mandatory is missing before drafting starts. To shred an RFP with Claude, give it every file in the pack as its own file, run the free Compliance Matrix Builder skill, read its list of anything it could not read, and check the result in both directions.

This guide is for proposal teams taking a solicitation apart. Tenders for document-shredding services are a different subject.

Shredding is the first working stage of how bid teams use Claude across an RFP, between the bid decision and the first draft. This guide covers what counts as a requirement, where requirements hide, six steps to shred with Claude, and the checks to run before the matrix reaches your writers.

What Is RFP Shredding?

RFP shredding is the first pass on a new tender: the team takes the solicitation apart into every discrete requirement, instruction and scored item, each in the buyer’s exact words with the place it came from.

The result is a requirements register, which becomes the compliance matrix and the submission checklist. For the columns a finished matrix carries, see what goes in a compliance matrix. In the bid management process, the shred sits after the decision to bid and before anyone is assigned a section to write.

Many teams already hand that first read to a general-purpose AI assistant, according to Stargazy’s 2026 RFP analysis survey, which it ran with a bid software vendor:

What Counts as One Requirement

Give every obligation someone could check on its own a row of its own:

• Bidder obligations: must, shall, will and “required” statements aimed at the bidder.
• Submission rules: instructions, formats, page limits, forms and attestations.
• Pass/fail items: anything that can disqualify the bid.
• Scored items: each evaluation criterion and the points it carries.
• Technical and security rows: each row of a requirements table and each security control.
• Price lines: each line the pricing form asks for.
• Q&A answers: each answer that adds or changes an obligation.

Split compound clauses that join obligations someone would check separately. The buyer’s own rights and background facts are context and stay out of the register.

Where Requirements Hide

Requirements rarely sit only in the body text. Look in each of these places:

• Body text: the sections a first read covers.
• Word tables: requirement matrices, service levels and response forms.
• Embedded files: a security workbook or pricing sheet embedded in the RFP’s Word file, which a text read shows only as an icon or an instruction to double-click.
• The Q&A or clarification log: many RFPs make the buyer’s answers binding, so one answer can add a requirement that appears nowhere else.
• Addenda: each can add, change or remove a requirement.
• The scoring table: evaluation criteria often sit in a separate table that a search for “must” never finds.

One miss can cost the bid. The Compliance Matrix Builder skill’s own guidance notes that evaluators have disqualified bids for missing a single mandatory submission instruction.

How to Shred an RFP With Claude, Step by Step

To build a compliance matrix from an RFP with Claude, decide to bid, collect every file (including anything embedded inside them), run the Compliance Matrix Builder skill, read what it says it received, then make the Q&A and compound clauses part of the register.

1. Decide to Bid Before the Full Shred

Run a go/no-go review first, testing only the mandatory items against what your company can deliver. The decision often turns on a handful of requirements, and a 400-row matrix is a big job to check for a bid you should not make.

The free AI Go/No-Go skill carries its own shredder for mandatory items and treats a deal-breaker as No-Go whatever the score. Give it the Q&A too, because a mandatory item can appear only in the buyer’s answers. See how the go/no-go skill is built.

2. Collect Every File and Save Embedded Objects Separately

Gather the main RFP, each appendix and form, pricing workbooks, the Q&A or clarification log and any addenda. Then open each Word file and look for embedded objects, such as an icon you are told to double-click.

Claude can only shred what reaches it. A security workbook embedded in a Word file can hold dozens of requirements, yet a text read shows only the instruction to open it. Open each object in Word and save it as its own file, and repeat this with every addendum.

Common mistake: assuming an uploaded Word file means Claude read everything inside it.

3. Install the Compliance Matrix Builder Skill

Install the free Compliance Matrix Builder skill, unlocked with an email address. It is a readable SKILL.md file, so you can check every instruction before you run it.

A skill applies the same method to every tender. Version 1.1 returns:

• A “Not read” list of every component it could not read, with a request for each missing file.
• A submission checklist ordered by deadline.
• The matrix: each requirement quoted in full with its source, category, priority and weight, plus blank Owner and Status columns.
• Separate rows for compound obligations, Q&A answers and scored sections.
• Context: evaluation criteria, key dates and observations on ambiguous or contradictory requirements, with an optional .xlsx export.

Run it in a Claude project for a few files, or in Claude Code for a folder. A long multi-file pack needs room to read every file, so don’t start one in a plain chat.

Using a prompt instead of the skill? Ask for the same things in one message:

Before extracting, list every component this tender refers to, including embedded files, the Q&A and named attachments, and mark each read, partly read or not available. Then list every requirement as its own row: quote it in full with no shortening, give its source section, and tag it mandatory, desirable or informational. Give each obligation in a compound clause its own row, add a row for every Q&A answer that adds or changes an obligation, and keep scored sections as rows.

4. Read What the Skill Says It Received

Before it extracts anything, the skill lists every component the tender refers to (embedded objects, the Q&A, named attachments), marks each read, partly read or not available, and asks whether to proceed.

That list is the cheapest place to catch a missing file. Compare it with your own list from step 2 before you reply “continue.”

5. Make the Q&A Part of the Register

Give every Q&A answer that adds an obligation its own row, citing the question number. Record every answer that changes a requirement on that requirement’s row.

An obligation left in a reference document never gets an owner. Version 1.1 of the skill makes these rows by instruction. With a prompt, ask for them explicitly, because a prompt can keep the Q&A word for word in a side table and still leave its obligations out of the rows.

6. Split Compound Clauses and Keep Scored Items as Rows

Give each obligation in a compound clause its own row. Two obligations in one row get one answer, and the second is easy to under-answer.

Keep evaluation criteria and their points as rows too, so a filter on mandatory items doesn’t drop them. Version 1.1 of the skill does both; with a prompt, ask for one row per obligation.

Check the Shred in Both Directions

Check an AI-built requirements list from both ends: rows against the source catch wrong or misquoted rows, and the source against the rows catches missing ones. Book a person’s time for both before the matrix reaches your writers.

From the rows, recheck the quotes, every number and date, any row shortened with an ellipsis, and each low-priority tag. From the source, confirm that every file, embedded object, appendix and Q&A answer has rows.

Where the Q&A and a tender document disagree, a person decides which one the bid follows, and asks the buyer if the question window is still open. With a prompt instead of the skill, ask Claude directly for its list of anything it could not read.

The checked register becomes a proposal compliance matrix your writers answer against. Later in the bid, the RFP Contradiction Checker skill is the gate before submission. For what Claude gets wrong in drafting, approved wording and long responses, see where Claude goes wrong beyond the shred.

A Shred on a Public Tender: What It Caught and Missed

We tested this method on State of Maine RFP# 202506081, a public state agency software RFP issued on July 15, 2025 and closed on August 8, 2025 (State of Maine RFP archive, read November 6, 2026), for a fictional bidder, with one run per step in a Claude Code session (Claude Opus 5.5). A hand-built answer key listed 457 requirements: 419 mandatory, 67 only in a security workbook embedded in the RFP’s Word file, and 44 only in the Q&A.

One run each: skill v1.1 on November 9, 2026; prompt 31 from the 72 Claude Prompts pack on November 6, 2026. No run invented a requirement.

• Without the workbook, every miss was in the workbook. Version 1.1 caught 390 of 390 requirements in the text it received, and its “Not read” list named the workbook first and asked for it.
• A prompt kept the Q&A but not its obligations. Prompt 31 kept all 191 Q&A answers word for word in a separate table, but turned only 7 of 44 Q&A-only obligations into rows.
• The first run showed what to fix. Version 1.0 of the skill found 385 and 452, losing requirements to quotes shortened with an ellipsis and leaving scored sections out of the rows. We changed the skill and re-ran both tests. The fix cost precision: more rows (482 and 511), duplicates up from 7 to 88 and from 17 to 60, and 91.7% and 92.6% of rows real bidder obligations, against 95.8% and 94.4% before.

The inputs decided the result, so supply every file, make the Q&A answers rows, and run the two-direction check. These are single runs with repeats pending, and the same team built the answer key and the skill.

Where AutoRFP.ai Fits Once the Requirements Are Listed

After the shred, every requirement needs an answer the company can stand behind. AutoRFP.ai drafts each requirement’s answer only from content the team has approved, shows the source behind it, and routes any requirement that content cannot support to a person.

As AI proposal software that answers each requirement from approved content, AutoRFP.ai starts from the buyer’s own files:

• Import: AutoRFP.ai imports the buyer’s Word files, Excel workbooks and PDFs, with each worksheet as its own import unit, so a security workbook saved as its own file comes in whole, and the team confirms the question mapping before anything is drafted from approved content.
• Review: each answer AutoRFP.ai drafts from approved content carries a Trust Score for how strongly its best source supported it, so reviewers start with the weakest, and any requirement the approved content cannot support is flagged for a person.
• Export: AutoRFP.ai’s exact-format export writes the approved, sourced answers back into the buyer’s own Excel or Word file; a PDF imports for answering and comes back as a Word document.

Claude stays in the workflow through AutoRFP.ai’s MCP connector: Claude can look up the team’s approved library with read scopes and, with full scopes, write to projects and content, acting as the signed-in user, with approvals staying in AutoRFP.ai’s workflow.

Teams comparing RFP shredding software can test any tool on a pack they have already answered, against a hand-checked list of its requirements. For whether Claude alone is enough or the whole response needs purpose-built software, weigh how Claude compares with purpose-built RFP software.

Where to Start on Your Next Tender

Before the next shred, open every file in the pack and write down what is embedded in it. That list is the first check on whatever Claude returns.

Then run the skill, read its list of what it received before you reply “continue,” and book a person’s time for the two-direction check.

Once the register is checked, the next stages are win themes and drafting in the buyer’s file, covered in the walkthrough of the rest of an RFP response in Claude.