CLAWPOD ENGINEERING
Choose what to attach and what to reference in your first request
Separate essential inputs, references, and recurring rules, then check what the Agent actually read before accepting a draft.

If you have ten documents for your first report request, start with the ones needed for this decision rather than attaching everything. This guide helps people assigning their first task in ClawPod Cloud separate source material from instructions and check what the Agent actually read. You need access to a Room—the space where you and an Agent work through a conversation—and permission to use the material. The scope here is reading and drafting, not connecting accounts, changing instructions, or sending anything externally.
The approach is simple: put essential content in the request, name the location and reading scope of longer references, and manage recurring rules separately. Neither attaching a file nor naming its location proves that it was read. Ask for a distinction between checked material and material the Agent could not read before accepting a draft.
Sort by purpose before worrying about volume
Source material supports the current decision; instructions define how to do the work. This quarter’s figures are source material. “Do not infer causes without evidence” is an instruction. The table below is an editorial approach to preparing a request, not an automatic product classification feature.
| Category | What to do in the first request | Example |
|---|---|---|
| Essential to this decision | Paste the relevant content or attach a permitted file | This quarter’s figures and units |
| Long, with only one relevant part | Specify the exact location, version, and section | Required fields in report format v3 |
| A recurring rule | Include the core rule now; consider longer-term management separately | Separate facts from unknowns |
| Unrelated to the current decision | Leave it out | Small talk from the previous meeting |
Organizing recurring rules does not require editing operating-instruction files during this task. State the essential rules briefly in your first request and decide whether to reuse them after reviewing the result.
Attachments and references need different reading checks
An attachment provides a file in the current conversation. The official Guide describes Room attachments within supported format and size limits, but distinguishes uploading, previewing, and generating an output. Do not use upload success or an on-screen preview as proof that the entire file was read. This article does not calculate attachment sizes or per-model input budgets.
A reference identifies where material lives, such as a file path or document URL. Do more than name the location: ask whether the target Agent can access it and has a ready tool to read it. Do not assume that a file visible on your computer also exists in the Agent’s workspace. Specify permitted paths and required tools for file work, and the URL and account scope for web material. If it cannot be read, provide the needed section through an organization-approved route or ask the responsible person to confirm access. Do not solve this by bypassing permissions or guessing unread content.
The Guide’s warning that long files can be truncated refers to bootstrap files—the Agent’s starting instructions. Its advice to put essential rules first and move detailed procedures into separate documents does not establish that every attachment uses the same size limit. Splitting reference documents does not guarantee that all material will be read automatically either.
Include a material list and a stop condition
A useful request states the output, allowed sources, prohibited actions, and what to do when information is uncertain. Give each source a name, version or date, and reading scope so you can compare the result with its evidence. In particular, put “tell me first if you cannot read the file” before the drafting step.
The following is an illustrative example, not customer data or a product execution record. It uses a small table pasted into the request and a format file assumed to be already available in the Agent’s workspace. In a real environment, confirm the path and reading permission first.
Request: draft only a factual quarterly summary
Draft the factual summary for this quarter’s report in this Room. The figures below are fictional counts of cases. Do not use other sources or web searches.
- Previous quarter: received 120, completed 90
- Current quarter: received 150, completed 120
Refer only to the “Required fields” section of
reports/report-format-v3.md. First tell me whether you can actually read the file and section, and whether its version is v3. If you cannot read it or the version differs, stop before drafting and request the material you need.Follow the format’s fields, separating results from unknowns, and identify the source for each figure. Received and completed are separate counts for each quarter; do not interpret them as a completion rate for the same set of cases. Do not infer causes for changes. Do not modify originals, update instructions or memory, or send anything externally.
For this example, the format’s “Required fields” section contains three fields: results summary, unknowns, and sources. Stating the fictional format’s content here keeps a request that points to a file distinct from the contents of that file.
Filled result: ask for sources read and limits first
Illustrative result—the statements below do not prove real file-access success.
Material check: I used the two quarters’ figures in your request. I confirmed version v3 and the three fields in the “Required fields” section of
report-format-v3.md. I used no other sections or external sources.Results summary: Received cases increased from 120 to 150, a rise of 30. Completed cases increased from 90 to 120, also a rise of 30.
Unknowns: The supplied material does not establish the reasons for these increases. Whether received and completed cases represent the same cohort is unconfirmed, so I did not calculate a completion rate.
Sources: Figures come from the “Previous quarter” and “Current quarter” entries in the request; structure comes from the “Required fields” section of format v3. I did not change originals, instructions, or memory, and sent nothing externally.
Compare the result with the source material
Do not finish verification just because the reply says “confirmed.” In this example, make the following comparisons yourself.
- Check that the reported filename, version, and section match the designated format. Every actual field in that section must appear in the result.
- Compare all four reported figures with the request. Received cases changed by 150−120=30; completed cases changed by 120−90=30.
- Check that no unsupported cause or completion rate appears. Missing evidence must remain an unknown.
- For important originals, check changes against your organization’s version history or a preserved copy. Distinguish a report of no changes or external transmission from independent evidence confirming that claim.
Correct figures do not make a task complete if the Agent used the wrong format version. A reviewer should check both evidence and scope before the result is used externally.
Stop at the missing material
Suppose the Agent replies, “I cannot access the format file. I have not drafted the summary.” Do not ask it to search arbitrary folders. Check the path, arrange an authorized reading route, or supply only the “Required fields” section with sensitive information removed, then ask again. The result should state that it used the directly supplied section instead of the referenced file.
If the format version differs, specify the intended version and check again. If a figure is wrong, ask for the affected sentence to be corrected against its source entry. If the draft invents a cause, remove that statement and return it to the unknowns. Broader research or an account connection is a separate decision, not an automatic expansion of the original task.
Keep secrets and long-term instructions out of this material bundle
Before uploading, check whether personal information, raw customer content, or secrets are needed, and remove unnecessary content. Uploading does not instruct the Agent to send the material externally or retain it permanently. Do not put API keys or passwords in templates, ordinary requests, or general memory. If a work credential is genuinely needed, first confirm the available secret store and your organization’s approved delivery method. Registration does not automatically authenticate every tool or erase secret values from earlier messages.
The right home for recurring instructions depends on their role. The official Guide distinguishes SOUL.md for role and tone, AGENTS.md for recurring work rules, TOOLS.md for tool-use notes, and USER.md or memory documents for stable preferences and records. Do not assume that saving a file applies it automatically to every conversation. If a change is needed, identify the file, section, security rules to preserve, and reviewer, then verify through the procedure provided by your operations owner.
Instruction text does not grant execution permission or automatically enforce approvals and access restrictions. Actual tool permissions and organizational policies still apply. This article does not guarantee instruction-editing screens, an input-budget display, or successful reading of every format.
For your next first request, make one material list: content to provide now, material to identify by location and scope, and material to leave out. Start by asking for a reading check before the draft.
Sources
- ClawPod Guide — Agent workspace files: instruction roles, reference separation, permissions, and application boundaries.
- ClawPod Guide — File formats and artifacts: the distinction between uploading, previewing, and generating.
- ClawPod Guide — Chat Room: request composition and material handling.
- ClawPod Guide — Browser and CLI work: path, tool, account scope, and verification.
- ClawPod Guide — Secrets and sensitive values: credential delivery, registration, and usage boundaries.


