This contents list supports the guide to context engineering in a real workflow, and uses the same fictional service credit case: a managed IT provider deciding whether an outage earns a contractual credit under the schedule version in force on the incident date.
The packet in assembly order
Order matters because each document depends on the one before it. The task statement defines what context is needed, the inventory names it, and the version register shows whether it survived.
| # | Document | What it shows | Who prepares it | Length |
|---|---|---|---|---|
| 1 | Cover sheet | Company size, systems, years covered, authorized contact | Introducer with the owner | One page |
| 2 | Task statement | The decision being tested, in one or two sentences | Service delivery lead | A paragraph |
| 3 | Context inventory | Documents a correct decision needs, and those it must ignore | Service delivery lead with legal | One table |
| 4 | Version register | Each policy revision with effective dates and where it is stored | Document control or IT | One table |
| 5 | Authority summary | Who could approve what, and the limits in force by year | Finance | Half a page |
| 6 | Outcome and dispute summary | Counts of decisions, credits, disputes and reversals per year | Billing or accounts receivable | One table |
| 7 | Acceptance check | The test that proves an answer right, written before outcomes are read | Service director | A paragraph |
| 8 | Rights and sensitivity notes | Customer data, contract terms limiting reuse, who must consent | Legal | Half a page |
What each document must say, and must not
1. Cover sheet. Peak full-time headcount (contractors excluded), the ticketing, billing and document systems in use with their years, the date of any migration, and the name of the person who can authorize a licensing discussion. No customer names.
2. Task statement. For the service credit case: "Decide whether a customer is owed an SLA credit and how much, citing the governing schedule." Keep it to one decision. A packet that tries to cover every ticket type is impossible to check.
3. Context inventory. Two lists. Required: the incident ticket, the credit request, the contract clause naming the governing version, the schedule version in force, the monthly fee and the approver authority table. Excluded: later schedule revisions, other customers' credit decisions and draft schedules never adopted. Anthropic frames good context engineering as finding the smallest set of high-signal information that makes the desired outcome most likely; the excluded list is where a packet shows it understands that.
4. Version register. One row per revision: revision number, effective from, effective to, storage location, and whether the file still exists. Microsoft notes that a SharePoint library saves only up to a set number of versions and deletes the oldest when the limit is reached, so the "still exists" column is not a formality.
5. Authority summary. The approval limits by role and year, for example a service director's goodwill limit. It describes the table; it does not attach approvals.
6. Outcome and dispute summary. Counts only: credit requests per year, approved, declined, disputed, reversed, and which system holds the dispute.
7. Acceptance check. Written before anyone looks at outcomes, so the test cannot be tuned to the answers. For this case: the cited schedule version equals the version in force on the date the contract names; the computed credit matches that version's formula; any goodwill credit is labeled separately and sits within the approver's limit.
8. Rights and sensitivity notes. Whether customer contracts restrict reuse of their records, what personal data appears in tickets, and whether a parent company must approve.
A filled-in task sheet for the service credit case (Illustrative)
Documents 2, 3, 6 and 7 are often combined into one task sheet. Here is the fictional provider's version, with counts the operations lead estimated in conversation.
| Field | Entry |
|---|---|
| Task | Decide SLA credit eligibility and amount, citing the governing schedule |
| Required context | Incident ticket, credit request, clause 9.2, schedule version in force, monthly fee, authority table |
| Excluded context | Later schedule revisions, other customers' decisions, unadopted drafts |
| Allowed actions | Approve contractual credit, decline, refer to service director, approve goodwill within limit |
| Acceptance check | Cited version matches incident date per clause 9.2; credit matches that version's formula; goodwill labeled and within limit |
| Volume and span | About 300 credit requests a year, six years of history, ticketing migration in 2021 |
| Known gaps | Schedule edited in place before 2023; disputes held in email |
The known-gaps row matters. Our evidence gap analysis for context engineering shows how to word those gaps so the review starts from an accurate picture.
How a reviewer uses the packet
The packet lets a reviewer decide whether past decisions can be turned into tasks with a checkable answer. micro1, a data company, describes on its blog (checked October 18, 2026) a review tool in which a reviewer inspects a model's full attempt, including tool calls and output, against the task's rubric. A packet whose acceptance check is concrete enough to grade that way is far easier to assess than one that says "correct decisions were made".
The reviewer reads the packet in order and stops at the first document that fails. If the version register shows most revisions lost, the acceptance check cannot be applied to those years, and the reviewer narrows the scope rather than rejecting the company. If the rights notes show a parent company controls the records, nothing else matters until that is settled.
What stays out until an agreement is signed
- Records themselves. No tickets, schedules, contracts or billing extracts. The packet describes them.
- Customer identities. Counts and categories only.
- Recreated notes. Writing reasons into old tickets now would corrupt the evidence. Describe the gap instead.
- Value estimates. Nobody in the packet guesses what the records are worth.
Owners preparing for the discussion can start with questions to ask before licensing company data. Contract-heavy review work is illustrated in the MCP M&A contract review example, and engineering sign-off trails in engineering QA and QC process data. For the first conversation with the owner, use the context engineering discovery questions.
The company still has to meet the basics: 50+ full-time employees at peak (contractors excluded), documented operating history, rights to license the records and an authorized contact. Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English. Whether any buyer wants the task, what de-identification is needed and what the company would be offered stay unknown until SourceX reviews the packet. Other record families are listed under data types.
Next step
Refer a company. No cap on rewards. Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.
If the owner can fill in the cover sheet and names a service lead who could prepare the rest, register as a partner, confirm who qualifies and read the terms before you make the introduction.