Implementation · Agent loading

How to load an agent into Claude, ChatGPT, or Gemini.

Every Valore AI agent ships as a portable Markdown file plus paired skills and templates. Loading it into your preferred reasoning model takes five steps. Worked through with the Credit Analyst + Investment Memo Drafter + UW Workbook as the example.

Model-agnostic — Claude, ChatGPT, Gemini · Portable files, no platform required · Human review required on every output

A. What you receive

Concrete file inventory per agent purchase

What the agent does

Reads your structured inputs, drafts the deliverable in the format the skill file specifies, surfaces missing information, flags assumptions out of band.
What the agent does NOT do

Make credit decisions. Send external communications. Replace counsel, tax, or insurance review. Guarantee accuracy. The QA checklist ships with each agent — use it.
What you receive in the file pack

Agent definition (.md) · paired skill files (.md, one per skill) · paired template files (.xlsx / .docx) · QA checklist (.md) · this how-to-load guide.
How files version

Each release of the agent + skill files ships with a version stamp in the header (e.g. credit-analyst-v1.0.md). You're responsible for tracking which version you ran a workflow against; reviewed outputs should reference the version in their metadata.

B. Five steps per platform

Pick the model your firm runs

Claude (Anthropic)

via Claude Projects

1
Open claude.ai and start a new Project (or open an existing CRE deal project).
2
In the Project settings, paste the entire content of Credit Analyst.md into the Custom Instructions field. This is the agent definition: system prompt, tool permissions, source-citation discipline, output format constraints.
3
Upload the supporting files to the Project's Knowledge: the Investment Memo Drafter skill file (Markdown), the populated UW Workbook (.xlsx), and any deal-specific inputs (sponsor brief, rent roll, T-12).
4
Start a new chat in the Project. The agent runs the moment you submit your first message. First message should be a structured task ("Draft the screening memo from the attached UW Workbook"), not "what can you do?"
5
Review the output against the QA checklist that ships with each agent. Save the reviewed output back into the Project Knowledge so the next call has context.

Recommended for sustained deal-team workflow — Project Knowledge persists across chats. Claude Sonnet 4.6 or Opus 4.7 recommended for credit-side work.

ChatGPT (OpenAI)

via Custom GPT or Project

1
Open chatgpt.com and go to Create a GPT in the sidebar (Plus / Team / Enterprise required).
2
In the GPT configurator, paste Credit Analyst.md into the Instructions field. Skip the "Conversation starters" and tool-toggle defaults; the agent definition file specifies its own.
3
Upload the Investment Memo Drafter skill file (Markdown) + the populated UW Workbook (.xlsx) into the GPT's Knowledge base.
4
Save the GPT (private to you or shared with your workspace). Start a new chat with the GPT; submit the structured task.
5
Review the output. Custom GPTs do not persist conversation context across chats by default — copy the reviewed output into the GPT's Knowledge if the next task should build on it.

Custom GPTs work well for repeatable workflows but lack Claude Projects' persistent conversation memory. For sustained deal-team work, Claude Projects is the better fit.

Gemini (Google)

via Gem

1
Open gemini.google.com and go to Gem Manager.
2
Create a new Gem; paste Credit Analyst.md into the Instructions field.
3
Attach the Investment Memo Drafter skill file + the populated UW Workbook (.xlsx) as Gem files.
4
Save the Gem. Open a new chat with the Gem; submit the structured task.
5
Review the output. Gem context behaves similarly to Custom GPTs — no automatic cross-chat memory; reviewed outputs need to be re-uploaded if subsequent calls should build on them.

Gemini 2.5 Pro recommended for credit-side work. Spreadsheet-handling is strong; PDF-handling is weaker than Claude on complex term-sheet documents.

C. Troubleshooting

First-time issues + the fix

Agent gives a generic answer instead of using the deal-specific files

Fix · Your prompt is too vague. Submit a structured task that explicitly references the file ("Draft the screening memo using the attached UW Workbook for [Deal Name]"). The agent will not infer that uploaded files are relevant.
Agent invents numbers not in the workbook

Fix · Claude and ChatGPT both occasionally fabricate when ambiguous. Add to your prompt: "Only cite numbers from the attached workbook. If a number isn't in the file, mark it as missing-information." Re-run.
Output format doesn't match the skill file's spec

Fix · The skill file may not be in the agent's context. Re-upload the skill file. Verify Claude Project Knowledge / ChatGPT GPT Knowledge / Gemini Gem files all show the skill file present.
Agent runs out of context window on long inputs

Fix · Use Claude Opus (1M context) or break the input into chunks. For Credit Analyst on a 12-tab UW Workbook, this almost never happens; for IR Associate on a 200-page LP letter history, it can.
Output looks plausible but is wrong in a way I can't articulate

Fix · This is the QA-checklist case. Every agent ships with a 5-10 item checklist for the kind of subtle wrongness it typically produces. Use it.

Next step

Build the stack the agent fits into.

An agent alone is a system prompt. The Valore architecture pairs each agent with skills, templates, and data products so the agent has something to operate on. Open the Company Builder and assemble the stack you’ll deploy — or jump to the template walkthroughs to see what each agent actually drives.

Open the Company Builder How to use templates Browse AI Team