Skip to content

contributor-initial-prompt.md

Latest commit

 

History

History
119 lines (87 loc) · 6.63 KB

contributor-initial-prompt.md

File metadata and controls

119 lines (87 loc) · 6.63 KB
Raw

Clone and set up the OneTyler Foundry knowledge-base repo so I can review agent transcripts.

git clone https://github.com/tyler-technologies/onetyler-foundry-team-agent-kb.git
cd onetyler-foundry-team-agent-kb

Then, in this order:

  1. Read CLAUDE.md in full, first, before doing anything else. It is your operating manual for this repo and it overrides your own defaults. Then read transcripts/ONBOARDING.md (my walkthrough as a reviewer) and transcripts/README.md (the process and every review field). Don't skim these — the whole workflow depends on details in them.

  2. Know what I may and may not change. Read .github/admin-only-paths.txt — that file is the boundary, so use it rather than your memory of this list.

    I can change: knowledge content in any Knowledge-<Domain>/ folder — the Conf-, Docusaurus-, FAQ-, Misc-, Training- and GitHub- files — plus my review verdicts under transcripts/. That's my job here; I know the subject matter.

    I cannot change: anything that decides which agent answers, or how the repo operates. README.md, team-config/, every Knowledge-*/_START_HERE.md, CLAUDE.md, transcripts/README.md, transcripts/ONBOARDING.md, scripts/, templates/, .github/, .gitignore, contributors.json.

    _START_HERE.md is the one that catches people out: it sits in a folder I can otherwise edit, but it carries cross-agent hand-off rules, so it's admin-only. Don't touch it.

    If you spot a genuine problem in any of those — a wrong command, a stale number, a contradiction — tell me and leave it alone. Don't fix it, and don't edit the CI check that enforces this. An admin decides. This matters more for you than for me: if you rewrite your own instructions mid-session you'll then follow the rewrite, and nobody reviewing my pull request can tell which rules you were actually working under.

    One thing no check can catch, so watch for it yourself: don't put routing advice inside a knowledge file. "For identity questions, use the Identity agent" in an FAQ is team-level routing in a file I'm allowed to edit. Flag it to me instead.

  3. Sync the reviewer list, so I can pick my own name as a reviewer:

    python3 scripts/sync_contributors.py --check   # exits 1 if it has drifted
python3 scripts/sync_contributors.py           # rebuild if it has

    This reads GitHub team membership using my own gh credentials. If it fails for lack of scope, tell me to run gh auth refresh -s read:org. If my GitHub username isn't in contributors.json afterwards, stop and tell me — I'm not on the team yet and I can't record a review until I am. Don't hand-edit the file to work around it; it's generated, the next sync overwrites it, and it grants no access anyway.

  4. Check whether I have a Foundry API key, since fetching transcripts needs one. Ask me whether I'm on macOS or Windows first — the setup differs.

    # macOS / Linux / Git Bash
test -n "$FOUNDRY_API_KEY" && echo "key set" || echo "NOT SET"
# PowerShell
if ($env:FOUNDRY_API_KEY) { "key set" } else { "NOT SET" }

    If I don't have one, tell me to create one in Foundry under Dev → API Keys, then give me the storage steps for my platform — the key goes in the OS credential store, not a file:

    macOS (Keychain):

    security add-generic-password -a "$USER" -s foundry-api-key -U -w
# then in ~/.zshrc:
export FOUNDRY_API_KEY=$(security find-generic-password -a "$USER" -s foundry-api-key -w)

    Windows (DPAPI — encrypted to my user on this machine, so a synced copy is useless):

    New-Item -ItemType Directory -Force "$env:LOCALAPPDATA\Foundry" | Out-Null
Read-Host -AsSecureString "Foundry API key" | ConvertFrom-SecureString |
  Set-Content "$env:LOCALAPPDATA\Foundry\key.dpapi"
# then in $PROFILE:
$sec = Get-Content "$env:LOCALAPPDATA\Foundry\key.dpapi" | ConvertTo-SecureString
$env:FOUNDRY_API_KEY = [Runtime.InteropServices.Marshal]::PtrToStringAuto(
  [Runtime.InteropServices.Marshal]::SecureStringToBSTR($sec))

    Either way I need a fresh terminal after editing my profile. If I insist on a file instead, tell me all three rules: outside the repo, outside any cloud-synced folder, and locked-down permissions. Watch for the trap — on macOS a checkout under ~/Library/CloudStorage/OneDrive-.../ makes "one level up" inside OneDrive; on Windows, Documents and Desktop are often redirected into OneDrive by policy. ~/.config/foundry/ and %LOCALAPPDATA%\Foundry\ are safe.

    Don't ask me to paste the key to you, and don't write it into any file in the repo. Keys are per-user and tenant-scoped, and I can hold 10. CLAUDE.md has the rest.

    Without a key you can still work from the transcripts already in the repo — say so and carry on rather than stopping.

  5. Start the transcript review UI in the background and confirm it responds:

    python3 scripts/review_server.py

    It serves http://127.0.0.1:7777 on loopback only. If that port is busy use --port 7778 and tell me the port you actually used.

  6. Show me the state of the queue — python3 scripts/review_status.py — and explain in your own words what I'm looking at: how many transcripts are waiting, and what the lifecycle states mean.

  7. Finish your reply with the review UI's URL, visually separated so I can't miss it, with the pending count and anything waiting specifically on me. Something like:

    ────────────────────────────────────────────
  Transcript review UI:  http://127.0.0.1:7777
  4 pending · 1 suggestion awaiting you
────────────────────────────────────────────

  8. Finally, read contributor-prompting-guide.md yourself and give me a short summary — five or six of the phrases I am most likely to need, in your own words. Do not paste the page back at me, and tell me plainly that it is a page to read rather than anything I paste into a terminal. I want to know what to say to you, not to be handed a document.

    In particular tell me the phrase that puts my work live, because nothing reaches my colleagues until I say it.

Then stop and wait. Don't start reviewing transcripts for me — reviewing is a human judgement and your verdict isn't mine. Once I've reviewed some, I'll ask you to process them.

If any step fails, say which one and what the error was rather than working around it.