# My New Agentic Campaign Finance Sidekick
Matt Hodges
2025-12-19

Claude [Agent
Skills](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview)
are pretty neat. They remind me of OpenAI’s [function
calling](https://platform.openai.com/docs/guides/function-calling). Both
are a way to tell the model to run some external process, often with the
goal of seeding the context window with deterministic data from outside
the LLM system. Function calling does this with a pile of JSON. Agent
Skills do it with a pile of Markdown.

> [!TIP]
>
> ### Now Works With Codex!
>
> OpenAI’s [Codex CLI](https://developers.openai.com/codex/cli/) now
> supports the open [Agent Skills standard](https://agentskills.io) so
> this now works with Claude Code or Codex!

[I like to do weird things with campaign finance
data](https://matthodges.com/posts/2025-08-02-language-model-embeddings-campaign-donors/),
and I thought it’d be nice to have an agentic sidekick to crawl around
FEC files with me. There’s a ton of prior art here, so it’s a great use
case for a Skill. The most relevant prior art is Derek Willis’s
[llm-fecfile](https://github.com/dwillis/llm-fecfile), which wraps Evan
Sonderegger’s venerable
[fecfile](https://github.com/esonderegger/fecfile), and is a plugin for
Simon Willison’s [LLM CLI tool](https://llm.datasette.io/en/stable/).

So I set out to make an Agent Skill. Except I didn’t have to do all that
much.

I popped open Claude Code (Opus 4.5) and asked for a plan:

> I have cloned the llm plugin llm-fecfile to @llm-fecfile/ . I would
> like to convert it to an Agent Skill. Let’s break it into tasks.

    ⏺ Here's the task breakdown for converting llm-fecfile to an Agent Skill:

      Tasks

      | # | Task                                                                       |
      |---|----------------------------------------------------------------------------|
      | 1 | Set up `.claude/skills/fec-filing/` with the required layout               |
      | 2 | Write SKILL.md file with YAML frontmatter and main instructions            |
      | 3 | Create FORMS.md with specific guidance for F1, F2, F3, F99 filings         |
      | 4 | Create SCHEDULES.md with field mappings for Schedules A, B, C, D, E        |
      | 5 | Create script to fetch and display FEC filings using the `fecfile` library |
      | 6 | Verify Claude can analyze an FEC filing using the new skill                |


      Key Conversion Notes

      - Plugin vs Skill: The LLM plugin uses a fragment loader (-f fec:ID). The skill will
      instead instruct Claude to use a Python script via Bash to fetch filings.
      - Documentation: The rich analysis instructions in llm_fecfile.py will become markdown
      reference docs that Claude reads.
      - Dependencies: The skill will require fecfile to be installed.

      Should I proceed with this plan, or would you like to adjust any tasks?

> Yes, let’s go.

From there, it began porting all of Derek’s hard work into the necessary
Markdown and script structure. Minutes later, I had
[agent-fecfile](https://github.com/hodgesmr/agent-fecfile) (formerly
called `claude-fecfile`) and you can have it too:

> [!WARNING]
>
> Installing someone else’s Agent Skill means you’re letting the agent
> run someone else’s **prompts** and **code**. I trust me. Do you trust
> me?

If you run this, the skill is now available globally in all your Claude
Code sessions:

``` bash
git clone git@github.com:hodgesmr/agent-fecfile.git /tmp/agent-fecfile
cp -R /tmp/agent-fecfile/skills/fecfile ~/.claude/skills/
```

Or if you run this, the skill is now available globally in all your
Codex sessions:

``` bash
git clone git@github.com:hodgesmr/agent-fecfile.git /tmp/agent-fecfile
cp -R /tmp/agent-fecfile/skills/fecfile ~/.codex/skills/
```

> [!NOTE]
>
> New install instructions are documented in the [agent-fecfile
> repository](https://github.com/hodgesmr/agent-fecfile).

Make sure you also have [uv](https://docs.astral.sh/uv/) installed.

And now you can get the agent to do things like this:

> Show me a table of the contribution counts and totals, by state, in
> fec filing 1896830

    ⏺ FEC Filing 1896830 - Contributions by State

      | State | Count |      Total | % of Total |
      |-------|-------|------------|------------|
      | UT    |    32 | $35,042.25 |      87.0% |
      | VA    |     2 |  $2,400.00 |       6.0% |
      | GA    |     1 |    $750.00 |       1.9% |
      | MI    |     1 |    $750.00 |       1.9% |
      | MD    |     2 |    $750.00 |       1.9% |
      | FL    |     1 |    $600.00 |       1.5% |
      | TOTAL |    39 | $40,292.25 |       100% |

      Utah dominates with 87% of the total contributions. The committee received from
      6 states total, with only 7 out-of-state contributions.

And things like this:

> What are the largest expenditures in filing 1896830?

    ⏺ FEC Filing 1896830 - Largest Expenditures

      | #   | Recipient                  |    Amount | Date       | Purpose          |
      |-----|----------------------------|-----------|------------|------------------|
      | 1   | Trump National Doral Hotel | $1,013.61 | 2025-05-19 | RNC training     |
      | 2   | Uber.com                   |    $19.77 | 2025-05-06 | Travel - Cabfare |
      |     | TOTAL (itemized)           | $1,033.38 |            |                  |

      Note

      Only 2 expenditures are itemized on Schedule B. The filing shows $21,283.49 in total
      disbursements, meaning most spending (~$20,250) was below the $200 itemization
      threshold and appears only in summary totals.

      The largest single itemized expense was for RNC training at Trump National Doral Hotel
      in May 2025.

> [!WARNING]
>
> FEC filings should be considered [untrusted
> content](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/).
> A malicious campaign sneaking prompt injections into the memo text
> field of their F99 is probably unlikely, but not impossible.

Agents!