How to Organise Client Documentation With AI | Knowledge Base

Use this process when a client sends a folder of HPWH evidence and it needs to be turned into a clean working document set.

The aim is to keep the original client files intact, then create two useful views:

all-documents/, sorted by document type such as 4234 reports, 5125 reports, drawings, manuals, data plates, and TRNSYS files.
all-models/, sorted by model or system ID so all evidence for one model can be reviewed together.

This process is designed for Claude or Codex. The AI does the repetitive file classification, but the engineer is still responsible for checking that the result makes sense.

When to use this

Use this workflow for client documentation folders in:

Energy AE\Files - Client_Documents

It is most useful for HPWH evidence packs that include files such as:

AS/NZS 2712 certificates
AS/NZS 4234 reports
AS/NZS 4692 reports
AS/NZS 5125.1 reports and test data
manuals
drawings
electrical safety evidence
data plates
product upload forms
TRNSYS .dck, .lst, and .out files

Do not use this process as a substitute for technical review. It only organises files and creates a starting index.

Folder setup

Start with a client folder like this:

CLIENT NAME/
  _document-sorter-kit/
  un-organised client files...

The package folders are the folders received from the client or prepared for a scheme. For example:

DAIKIN AUSTRALIA PTY LTD/
  _document-sorter-kit/
  Daikin - ESS/
  Daikin - VEU/

Ask Claude or Codex to sort the folder. The AI should install the sorter workflow and move the package folders into raw/processed/, so the client root ends up like this:

CLIENT NAME/
  all-documents/
  all-models/
  docs/
  raw/
    unprocessed/
    processed/
      general client files
  tools/
  _document-sorter-kit/
  _system/
  CLAUDE.md

raw/processed/ is the immutable source area for existing material. Do not delete or rename source files there unless there is a clear reason.

Use raw/unprocessed/ only for new files received later, after the first sorting pass has already been completed.

Prompt to use

Use a short prompt. The workflow should be clear from the folder structure and the _document-sorter-kit.

Example prompt:

Sort this client documentation folder using the _document-sorter-kit.

Move any package folders currently beside _document-sorter-kit into raw/processed/.
Install the generic sorter files at the client root.
Configure the client brand and model prefixes from the file names.
Run the sorter.
Update docs/index.md and docs/log.md.
Tell me any questions or review items.

If the folder structure is not clear to the AI, stop and improve the kit or this instruction note. Do not rely on a long one-off prompt that only works once.

What the AI should do

The AI should:

Read _document-sorter-kit/README.md and the generic instructions.
Copy the reusable workflow files from _document-sorter-kit/generic/ into the client root.
Create the starter folders from _document-sorter-kit/client-starter/ if they do not already exist.
Move existing package folders from the client root into raw/processed/.
Configure _system/client-config.json with the client brand and model prefixes.
Run a dry run of tools/sort_client_documents.py.
Review _system/processing-summary.md, _system/review-needed.csv, and _system/copy-errors.csv.
Adjust obvious document classification rules if files have been placed in misc-review incorrectly.
Run the sorter for real.
Update docs/index.md and docs/log.md.

After confirming an incremental batch is handled, move those source files from raw/unprocessed/ into raw/processed/.

What to check after sorting

Check the generated summary first:

_system/processing-summary.md

The summary should show:

how many raw files were scanned
how many files were selected for all-documents/
how many files were selected for all-models/
how many copy errors occurred
how many review-needed rows remain

Then check:

_system/review-needed.csv
_system/copy-errors.csv
docs/index.md
a few representative folders in all-models/
a few representative folders in all-documents/

The review queue does not always need to be zero, but every row should be understood. A file in misc-review may be fine if the evidence type is genuinely unclear. It should not contain obvious items such as TC5.csv test data or a manual that the sorter failed to classify.

What good output looks like

Good output has:

no client package folders left loose at the client root
original source folders preserved under raw/processed/
one document-type view in all-documents/
one model-level view in all-models/
_system/sort-manifest.csv showing where every source file was classified
_system/review-needed.csv listing only genuine uncertainties
_system/copy-errors.csv empty
docs/index.md summarising the known models and evidence gaps
docs/log.md recording what was done

The AI should not delete raw client files. It may delete stale generated files from all-documents/ or all-models/ if it verifies they are generated outputs and are no longer present in the manifest.

How to review the model index

The model index is a working aid, not a source of truth by itself.

Use docs/index.md to see which model IDs were found and what evidence is available. Treat fields such as product family, system type, compressor speed, Wi-Fi, set point, deadband, and approval status as TBC unless the AI has cited a source document or the engineer has checked the source directly.

If the model index includes evidence-only model IDs, reconcile them before relying on the folder for a submission. Evidence-only IDs may mean:

the model prefix was not configured correctly
the source document uses a component model rather than a sale model
one system has several related model names
a certificate or report covers a model family

Common fixes

If many files are _UNMATCHED, update _system/client-config.json with the correct brand and model prefixes, then rerun the sorter.

If obvious document types are going to misc-review, update tools/sort_client_documents.py with a narrow rule. Keep the rule specific to the evidence type. Do not create broad rules that hide uncertainty.

If duplicate folders appear, check whether the package folder was accidentally copied instead of moved into raw/processed/.

If copy errors appear, check whether files are open in Excel, Word, Acrobat, or OneDrive. Close the file and rerun the sorter.

If the output looks wrong, do not manually rearrange all-documents/ and all-models/ first. Fix the sorter or config, then rerun it so the manifest remains reliable.

Responsibilities

The AI is responsible for organising the folder, creating the manifest, and identifying review items.

The engineer is responsible for checking that:

the model names make sense
the important evidence is present
the latest or correct report has been selected
the index does not invent technical details
the folder is suitable for the next project step

When the folder is ready, use all-models/ for model-level review and all-documents/ for document-type checks.