Skip to content

CLAUDE.md

Latest commit

 

History

History
105 lines (76 loc) · 5.13 KB

CLAUDE.md

File metadata and controls

105 lines (76 loc) · 5.13 KB
Raw

PL Infra AI Prompt Library — Project Instructions

What this project is

A shared AI prompt library for PL Infra teams (legal, finance, capital, operations, product, research, comms). The library lives in preview/prompts.json and is rendered as a static GitHub Pages site. Users browse prompts, copy them with one click, and submit new ones via a form that opens a GitHub issue — a workflow then turns the issue into a PR.

The source of truth for all prompts is preview/prompts.json.

Repo structure

Path Purpose
preview/prompts.json Prompt library data — the source of truth; the only file to edit to add, update, or deprecate prompts
preview/prompts.js Auto-generated mirror of prompts.json (window.PROMPTS) so the site runs over file://. Regenerate whenever prompts.json changes (see preview/README.md)
preview/index.html The library app — browse/filter/search/sort, upvote, in-page Prompting Guide tab, add-prompt modal, dark mode
preview/reference.html Redirect → index.html#guide (the guide is now in-page). Old standalone page preserved as reference.html.bak
preview/add-prompt.html Submission form (opens a GitHub issue → PR)
preview/app.js · add-prompt.js · styles.css Site behavior and styling. app.js reads window.PROMPTS
.github/workflows/prompt-from-issue.yml Turns add-prompt issues into PRs
.github/ISSUE_TEMPLATE/add-prompt.md Issue template for manual submission
CONTRIBUTING.md Full prompt entry schema and contribution workflow
sensitivity-framework.md L1–L4 data sensitivity tiers (full detail)
best-practices.md Prompting techniques — the Golden Formula, power-user tips, etc.
prompt-library-template.md Markdown template for teams that prefer it over the hosted site

examples/ is deprecated — ignore it.

How to help with common tasks

Adding or updating a prompt

When asked to add a prompt, always produce a valid JSON entry for preview/prompts.json. Use this schema:

{
  "title": "Short, descriptive name",
  "category": "Legal | Finance | Operations | Product | Research | Comms | Cross-functional",
  "use_case": "One or two sentences on when to use it.",
  "intended_audience": "Who this is for.",
  "recommended_model": "e.g., Claude Opus for deep analysis; Sonnet for drafting",
  "inputs_required": "What the user must supply before running.",
  "prompt_text": "Verbatim prompt text. Use {{placeholders}} for variable inputs.",
  "example_output": "3–5 line abbreviated sample showing what good output looks like.",
  "author": "Name / team",
  "last_updated": "YYYY-MM-DD",
  "status": "Draft",
  "sensitivity": "L1 | L2 | L3 | L4",
  "tags": ["tag1", "tag2"],
  "notes": "Optional caveats, gotchas, or related prompts."
}
  • Always set status to "Draft" on first add. It becomes "Approved" only after a second person has tested it.
  • Infer the sensitivity tier from the inputs the prompt is designed to consume (not the outputs). When uncertain, go one tier higher.
  • Use {{double_braces}} for placeholders — not angle brackets or ALL_CAPS.
  • Example outputs should show shape, not a full run — 3–5 lines is enough.
  • Never include real confidential data in examples. Use synthetic or redacted placeholders.

Deprecating a prompt

Set status to "Deprecated" in prompts.json. Leave the entry in the file — it becomes institutional memory and the site filters it to an archive view.

Checking or assigning a sensitivity tier

Reference this table. The rule is: tier is set by the inputs, not the outputs.

Tier Data type Tools allowed
L1 Public — open web, marketing, published regs Any
L2 Internal — working docs, OKR notes, anonymized data Enterprise-approved tools only
L3 Confidential — named third parties, unredacted contracts, financials, PII Enterprise-tier with data controls; DRI sign-off
L4 Restricted — privileged comms, M&A terms, gov IDs, API keys Human review by default; AI use requires explicit senior legal sign-off

When in doubt, go one tier higher and note the reasoning in the notes field.

Deploying the site

  1. Push to GitHub.
  2. Settings → Pages → Source: branch main, folder /preview.
  3. Update GITHUB_REPO at the top of preview/add-prompt.js to the live repo URL.
  4. Settings → Actions → General → enable read/write permissions and allow Actions to create PRs.

Running locally

cd preview
python3 -m http.server 8000
# open http://localhost:8000

Key principles to apply

The Golden Prompting Formula: ROLE + CONTEXT + TASK + FORMAT = Great Output.

When reviewing or improving prompts in the library, check that the prompt text includes all four components. If one is missing, flag it or suggest an addition.

When writing prompt text, use precise verbs: draft, summarize, extract, compare, flag, classify. Avoid vague asks.

When the user asks for prompting advice, draw from best-practices.md — especially the six power-user techniques and the "Always avoid" list.