benji/apes
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

init: README, CLAUDE.md, and claude skills (critic, gcloud, ax)

Browse Source 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
limiteinductive
2026-03-29 18:02:14 +02:00
commit 122290bae9
6 changed files with 383 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
.claude/settings.json Normal file
View File

@@ -0,0 +1,5 @@
{
"permissions": {
"dangerouslySkipPermissions": true
}
}

113
.claude/skills/apes/ax/SKILL.md Normal file
View File

@@ -0,0 +1,113 @@
---
name: ax
description: Audit agent-facing docs, hooks, skills, and config for the apes platform against AX principles. Use when agent behavior is wrong due to missing/unclear docs, poor ergonomics, or misconfigured automation.
argument-hint: "[problem description]"
disable-model-invocation: true
---
# AX — Agent Experience Audit
Audit the apes project's Claude Code configuration — CLAUDE.md, hooks, skills, rules, permissions — against AX principles. For each finding, recommend the right mechanism to fix it.
## Arguments
- `$ARGUMENTS` — description of the AX problem (e.g., "agents keep deploying to wrong project"). If empty, run a general audit.
## Workflow
### Phase 1: AUDIT — Discover and score
#### 1a. Establish ground truth
Derive canonical workflows from:
- `docker-compose.yml` files on VMs (SSH to check)
- Any `Makefile`, `package.json`, `pyproject.toml` in repo
- Deployment scripts, CI pipelines
- GCP project config (`apes-platform`)
Ground truth is authoritative. If docs and automation disagree, fix docs.
#### 1b. Inventory agent-facing surfaces
Discover all Claude Code configuration:
**Documentation:** `CLAUDE.md`, `.claude/rules/*.md`, `README.md`
**Automation:** `.claude/settings.json`, hooks
**Skills:** `.claude/skills/*/SKILL.md`
**Commands:** `.claude/commands/*.md`
**Agents:** `.claude/agents/*.md`
**Memory:** `~/.claude/projects/*/memory/MEMORY.md`
If `$ARGUMENTS` is provided, focus on relevant surfaces.
#### 1c. Score against AX principles
| # | Principle | FAIL when... |
|---|-----------|--------------|
| 1 | Explicitness over convention | A non-standard workflow isn't called out explicitly |
| 2 | Fail fast with clear recovery | Errors lack concrete fix commands |
| 3 | Minimize context rot | CLAUDE.md adds tokens that don't earn their keep |
| 4 | Structured over unstructured | Important info buried in prose instead of tables/code blocks |
| 5 | Consistent patterns | Naming or formatting conventions shift across docs |
| 6 | Complete context at point of need | Critical commands missing where they're needed |
| 7 | Guard rails over documentation | Says "don't do X" but X would succeed — a hook or permission would be better |
| 8 | Single source of truth | Same info maintained in multiple places, or docs diverge from reality |
**Apes-specific checks:**
- GCP project/region/zone correct everywhere?
- Docker Compose configs on VMs match what docs describe?
- DNS records match what's deployed?
- No SaaS dependencies crept in?
### Phase 2: PROPOSE — Select mechanism and draft fixes
For each WARN or FAIL, select the right Claude Code mechanism:
| If the finding is... | Use this mechanism |
|---|---|
| Block forbidden actions | **PreToolUse hook** |
| Dangerous command that should never run | **Permission deny rule** |
| Auto-format/lint/test after edits | **PostToolUse hook** |
| File-type-specific convention | **`.claude/rules/*.md`** with `paths` frontmatter |
| Repeatable workflow or reference | **Skill** |
| Complex task needing isolation | **Subagent** |
| Critical context surviving compaction | **CLAUDE.md** |
| Universal project convention | **CLAUDE.md** (keep <200 lines) |
Each fix must include:
- Which principle it addresses
- The selected mechanism and why
- Exact implementation (file path + content)
### Phase 3: REPORT
```
# AX Audit Report — apes
**Surfaces audited:** <count>
## Scorecard
| # | Principle | Rating | Detail |
|---|-----------|--------|--------|
| 1-8 | ... | PASS/WARN/FAIL | ... |
## Findings
| Surface | Issues | Recommended mechanism |
|---------|--------|----------------------|
| ... | ... | ... |
## Recommendations
For each:
- Principle addressed
- Mechanism type
- Exact implementation (file + content)
```
## Constraints
- This skill is **read-only** — it never modifies files, only reports
- Apes-specific: verify no SaaS dependencies in recommendations
- Verify GCP infra state via SSH before reporting on deployed services

66
.claude/skills/apes/critic/SKILL.md Normal file
View File

@@ -0,0 +1,66 @@
---
name: critic
description: Stress-test research hypotheses, architecture decisions, and vibecoded implementations with adversarial-but-fair critique. Returns structured JSON verdicts. Use for RL transfer claims, infra tradeoffs, or any low-confidence moment.
---
# Critic
Use this skill when the job is to make reasoning stronger, not to keep the conversation comfortable.
## Good fits
- RL transfer hypothesis validation — "will training on Go actually help with planning benchmarks?"
- architecture tradeoffs — self-hosted vs managed, monolith vs services
- vibecoded implementation review — "this works but was generated fast, is it sound?"
- research design — experimental methodology, benchmark selection, control groups
- infra decisions — GCP resource sizing, networking, security posture
- **ad-hoc low-confidence moments**: code behaving unexpectedly, ambiguous requirements, multiple valid approaches
## Do not use for
- routine implementation work
- simple factual lookup
- emotionally sensitive moments where critique is not the task
## Output contract
The critic always returns a JSON object as the first block in its response:
```json
{
"verdict": "proceed | hold | flag | reopen",
"confidence": 0.0,
"breakpoints": ["issue 1", "issue 2"],
"survives": ["strength 1", "strength 2"],
"recommendation": "one-line action"
}
```
Verdicts:
- **proceed** — no blocking issues
- **hold** — do not proceed until breakpoints resolved
- **flag** — notable concerns but non-blocking
- **reopen** — fundamentally flawed, needs rework
- **error** — critic could not complete (missing files, insufficient context)
Optional prose narrative follows after a blank line.
## Operating contract
- Be direct, not theatrical.
- Critique claims, assumptions, and incentives, not the person.
- If you agree, add independent reasons rather than echoing.
- If you disagree, say so plainly and explain why.
- Steelman before you attack. Do not swat at straw men.
- Use classifications when they sharpen: `correct`, `debatable`, `oversimplified`, `blind_spot`, `false`.
- For research claims, demand evidence or explicit acknowledgment of speculation.
- For vibecoded implementations, focus on correctness and security over style.
## Research-specific checks
When critiquing RL transfer hypotheses or experimental design:
- Is the hypothesis falsifiable?
- Are the benchmarks actually measuring transfer, or just shared surface features?
- Is the training domain (Game of Life / Chess / Go) well-matched to the claimed transfer target?
- Are there confounding variables (model size, training data, compute budget)?
- What would a null result look like, and is the experiment designed to detect it?

94
.claude/skills/gcloud/SKILL.md Normal file
View File

@@ -0,0 +1,94 @@
# gcloud Skill
Common GCP patterns for the apes platform. All commands invoke gcloud/kubectl/docker directly via Bash.
**Project:** `apes-platform`
**Region:** `europe-west1`
**Zone:** `europe-west1-b`
## Current Infrastructure
| Service | Host | VM | IP |
|---------|------|----|----|
| Gitea | git.unslope.com | gitea-vm | 34.78.255.104 |
| Chat (planned) | apes.unslope.com | TBD | TBD |
## SSH into VMs
```bash
# Gitea VM
gcloud compute ssh gitea-vm --zone=europe-west1-b --project=apes-platform
# Run a command remotely
gcloud compute ssh gitea-vm --zone=europe-west1-b --project=apes-platform --command="sudo docker ps"
```
## Docker Compose on VMs
```bash
# Restart a service
gcloud compute ssh <vm> --zone=europe-west1-b --project=apes-platform \
--command="sudo bash -c 'cd /opt/<service> && docker compose restart <container>'"
# View logs
gcloud compute ssh <vm> --zone=europe-west1-b --project=apes-platform \
--command="sudo docker logs <container> --tail 50"
# Full redeploy
gcloud compute ssh <vm> --zone=europe-west1-b --project=apes-platform \
--command="sudo bash -c 'cd /opt/<service> && docker compose pull && docker compose up -d'"
```
## Static IPs & DNS
```bash
# Reserve a new static IP
gcloud compute addresses create <name> --region=europe-west1 --project=apes-platform
# Get IP value
gcloud compute addresses describe <name> --region=europe-west1 --project=apes-platform --format='value(address)'
# DNS: add A record at Namecheap (Advanced DNS tab) pointing subdomain to IP
```
## Firewall Rules
```bash
# List rules
gcloud compute firewall-rules list --project=apes-platform
# Open a port
gcloud compute firewall-rules create <name> --allow=tcp:<port> --target-tags=web-server --project=apes-platform
```
## New VM Pattern
```bash
gcloud compute instances create <name> \
--project=apes-platform \
--zone=europe-west1-b \
--machine-type=e2-small \
--image-family=debian-12 \
--image-project=debian-cloud \
--boot-disk-size=20GB \
--tags=web-server \
--address=<static-ip-name> \
--metadata-from-file=startup-script=<script-path>
```
## IAM
```bash
gcloud auth list
gcloud projects get-iam-policy apes-platform --format=json
```
## Troubleshooting
| Error | Fix |
|-------|-----|
| VM SSH timeout | Check firewall: `gcloud compute firewall-rules list --project=apes-platform` |
| Docker not running | SSH in, run `sudo systemctl start docker` |
| Caddy cert failed | Check DNS propagation: `dig @dns1.registrar-servers.com <domain> A +short` |
| Container not starting | Check logs: `sudo docker logs <container> --tail 50` |
| DNS not resolving | Flush local cache: `sudo dscacheutil -flushcache && sudo killall -HUP mDNSResponder` |

54
CLAUDE.md Normal file
View File

@@ -0,0 +1,54 @@
# CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
## What This Is
`apes` is a post-singularity research platform. No SaaS. Everything vibecoded. Self-hosted on GCP.
**Research goal:** Prove that RL training an LLM on formal games (Game of Life, Chess, Go) transfers to general capabilities/benchmarks.
**Team:** Benji and Neeraj (the apes). Claude Code agents are first-class collaborators.
## Infrastructure
| Service | URL | VM | Zone |
|---------|-----|----|----|
| Gitea | git.unslope.com | gitea-vm | europe-west1-b |
| Chat (planned) | apes.unslope.com | TBD | europe-west1-b |
**GCP project:** `apes-platform`
**Region:** `europe-west1`
**DNS:** Namecheap (Advanced DNS tab for A records)
## Non-Negotiable Rules
- **No SaaS.** If we can't self-host it, we don't use it.
- **Vibecoded.** Humans direct, agents build. Move fast, verify correctness.
- **GCP project is `apes-platform`.** Always pass `--project=apes-platform`.
- **Region is `europe-west1`.** Zone `europe-west1-b` unless there's a reason to change.
## Route By Task
| Need | Load |
|------|------|
| GCP commands | `/apes:gcloud` skill |
| Stress-test a decision | `/apes:critic` skill |
| Audit agent config quality | `/apes:ax` skill |
## Deployment Pattern
All services run as Docker Compose on GCP Compute Engine VMs behind Caddy (auto HTTPS via Let's Encrypt).
```bash
# SSH into a VM
gcloud compute ssh <vm> --zone=europe-west1-b --project=apes-platform
# Manage services
sudo bash -c 'cd /opt/<service> && docker compose up -d'
sudo docker logs <container> --tail 50
```
## Critic Reflex
When something is surprising, contradictory, or your confidence is low, use the `/apes:critic` skill before proceeding. Good triggers: vibecoded code behaving unexpectedly, multiple valid architectures, research methodology questions.

51
README.md Normal file
View File

@@ -0,0 +1,51 @@
# 🐒 apes
**post-singularity research platform — vibecoded, self-hosted, no SaaS**
> the apes are the humans. everything else is smarter than us now. might as well own it.
## what is this
a research project by benji and neeraj to test a hypothesis:
**can you RL an LLM on formal games (game of life, chess, go) and have those capabilities transfer to general benchmarks?**
the twist: this entire project is vibecoded. no SaaS. no managed services. just apes prompting agents to build everything from scratch on GCP.
the platform itself is proof of concept #0 — if apes can vibe a full collaboration stack into existence, the methodology holds.
## live at
[apes.unslope.com](https://apes.unslope.com)
## phase 1: the colony
a self-hosted slack clone deployed on GCP where apes and claude code agents talk together as peers.
- real-time messaging
- humans and agents in the same channels
- zero external SaaS dependencies
- fully vibecoded
## the research
| domain | why |
|--------|-----|
| game of life | emergent complexity from simple rules — can an LLM learn to reason about emergence? |
| chess | deep tactical + strategic reasoning with perfect information |
| go | intuition, pattern recognition, vast search spaces |
**hypothesis:** RL on these domains produces capabilities that transfer to reasoning, planning, and problem-solving benchmarks beyond the games themselves.
## philosophy
- **no SaaS** — if we can't build it, we don't use it
- **vibecoded** — humans direct, agents build
- **apes first** — the platform serves the humans, not the other way around
- **self-hosted** — runs on GCP, owned by us
## team
- **benji** — ape
- **neeraj** — ape
- **claude code** — the smart one