Menu
Self-improving skill loop. Analyzes the current conversation — what worked, what didn't — and patches ae skills accordingly. Run at the end of a session.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
End-to-end validate a `text.ara.so/backend` (website-agent) fix against a real Railway preview environment BEFORE merging to main. Pushes the current branch, deploys the SHA to the long-lived `preview` Railway env, confirms the new `git_sha` is live via `/healthz`, then sends a focused signed webhook replay (single-turn or short multi-turn) to a reserved 555-01XX test phone, and polls Braintrust for the root span's `outcome` / rounds / tools / reply. Use whenever you change anything in `backend/src/` (tool handlers, system prompt, builder loop, etc.) and want production-fidelity proof — not just unit tests — before opening the PR. Invoked as `/replay-railway-preview` or `/replay railway preview`.
Ara's secrets convention — all runtime credentials live in Infisical (project "Ara-passwords"), one folder per active GitHub repo (`/ara-engineer`, `/text-ara-so`). The old megarepo lives at `/legacy-ara-megarepo` tagged `legacy` — reference only. Never ask the user to paste keys; never commit .env; never build a custom vault.
Align the website-agent (system prompt + routing) with successful on-Ara website building; analyze Braintrust traces for derailment (wrong phase, local-dev tutoring, missing deploy, paywall/connect confusion), patch `text.ara.so/backend` (primarily `system-prompt.ts`), then verify in a tight loop—`bt` / Braintrust evals / `bun run e2e` (replay + live)—and report a before vs after fit table. Invoked as `/align`, `/align <trace url>`, `/align users …` (pair with `/trace` to gather traces). Companion to `/trace`.
Ara agent-trace debugging — inspect Braintrust traces for the website-agent (TS/Bun, Cerebras, Vercel AI SDK v6). Invoked as `/trace recent`, `/trace turn <turn_id>`, `/trace convo <chat_id>`, `/trace user <phone>`, `/trace top users` (aggregate: who messaged most in a time window — one latest-turn link per sender, table below), `/trace tool <name>`, `/trace span <id>`, `/trace <url>`, `/trace test` (run canonical e2e via `text.ara.so/backend/scripts/e2e.ts` (`bun run e2e --target=local --scenarios=<name>`)), `/trace score`, `/trace online`, or `/trace grow`. **Use `/align`** to align system prompt and gates to traces and report before/after tables (`skills/align/SKILL.md`). **When the user asks for users’ traces / links:** give **one permalink per user** — the **latest** `webhook.inbound` root only. **When they ask for top / busiest users over a time range:** use the **top users table** format (Msgs, User, single review link) — no per-thread “all turns” column unless they ask.
Deeply research a technology already in the stack. Pulls latest official docs, changelogs, release notes, production recipes, and pitfalls via heavy web search — then maps findings onto the exact usage pattern in this repo to surface underused features and confirm best practices. Use whenever someone asks "are we using X right?", "what's the best way to use X for Y?", or "what did X ship recently that we should adopt?"
Braintrust `bt` CLI for Ara — every agent-facing service emits spans to the `Ara` project (org `Aradotso`). Use for inspecting logs/traces/prompts/evals, authenticating, SQL over spans, and wrapping new code with tracing. For the Ara-specific span shape and debug recipes (`/trace recent`, `/trace turn`, `/trace user`, etc.) see the companion `/trace` skill.
| name | improve |
| version | 1.0.1 |
| description | Self-improving skill loop. Analyzes the current conversation — what worked, what didn't — and patches ae skills accordingly. Run at the end of a session. |
| allowed-tools | ["Bash","Read","Edit","Write","Grep","Glob"] |
Analyze the conversation, extract learnings, patch skills. Manually triggered at session end.
All aracli changes live in ~/github/ara.engineer/ — the GitHub repo. Never edit ~/.aracli/ directly; that is a local install cache that gets overwritten on aracli update.
| What to change | Where to edit |
|---|---|
Skills (/demo, /ready, etc.) | ~/github/ara.engineer/cli/skills/<name>/SKILL.md |
CLI commands (ae url, ae pr, etc.) | ~/github/ara.engineer/cli/src/commands/<name>.ts |
| CLI entry + command registration | ~/github/ara.engineer/cli/src/index.ts |
Build three lists from everything that happened:
What worked — commands/approaches that ran clean, user confirmed without pushback What didn't — failures, corrections ("no not that"), missing context, wrong assumptions What was slow — redundant steps, info that had to be looked up repeatedly
Present these to the user. Get confirmation before making changes.
ls ~/github/ara.engineer/cli/skills/*/SKILL.md | sed 's|.*/skills/||;s|/SKILL.md||' | sort
For each learning, identify which skill to patch and what type of fix:
| Fix type | Example |
|---|---|
| Missing step | Skill forgot to check X before Y |
| Wrong assumption | Path/config that doesn't exist |
| Better default | Flag should be different |
| New pattern | Workflow that should be codified |
| Dead instruction | References something renamed/removed |
| New skill needed | Repeated workflow with no skill |
| New CLI command | Repeated ae invocation worth a native command |
For each approved change:
Skills:
~/github/ara.engineer/cli/skills/<name>/SKILL.md1.0.0 → 1.0.1)~/github/ara.engineer/cli/skills/<name>/SKILL.mdCLI commands (new or modified):
~/github/ara.engineer/cli/src/commands/<name>.ts~/github/ara.engineer/cli/src/index.ts (import + add to coreCommands)~/.aracli/cli/ — that is a generated install, not sourceRules:
cd ~/github/ara.engineer
git add cli/
git commit -m "improve: patch skills/cli from session learnings"
git push origin main
/improve complete
Updated: axiom (v1.0.0 → v1.0.1), ready (v1.0.0 → v1.0.1)
Created: —
Skipped: 1 learning (no skill match)