Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/drubinstein/terry/andrejnpx skills add drubinstein/terry --skill andrejgit clone --depth 1 https://github.com/drubinstein/terryWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00102 | $0.03804 |
| Opus 5 | $0.00051 | $0.01902 |
| Sonnet 5 | $0.00020 | $0.00761 |
| Haiku 4.5 | $0.00010 | $0.00380 |
Grade C, and why
andrej scanned grade C with 2 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Downloads and executes remote codehighSupply chain
curl | sh runs whatever the server returns today, which is not necessarily what it returned when this was reviewed.
curl -s http://127.0.0.1:8765/state | python3 -m json.tool # live state Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
run), babysitting a long agent run (signal = `curl /state`; metric = checkpoint How it starts
The opening of the file, as written. The whole thing — 284 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Andrej — the observe→analyze→act iteration loop
Andrej is the empirical decision-discipline you run inside every iteration when you are driving a live system toward a goal — a fix, a perf win, a converged run, a root cause. Instead of guessing a change from intuition or a sparse log, you gather ground-truth signal, interpret it against a mental model, make exactly ONE bounded change, measure a scalar metric, and keep-or- revert — then write down what happened and loop. The data picks the next action; you never guess.
Like its sibling terry, the skill carries a human name: Andrej nods to Andrej Karpathy and the autonomous edit → run a fixed budget → measure a scalar → commit-or-revert (via git) agent loop it generalizes (see References). This skill is that loop made explicit and domain-agnostic: it applies to CI triage, performance tuning, flaky-test hunts, hyperparameter search, incident response, and babysitting any long-running autonomous run.
It is the generalization of five primitives wired into one loop:
- signal — ground-truth observation: a trace, a metric stream, a status endpoint, a repro, a probe of the live code path (NOT a buffered log grep)
- scalar metric — the one number that mechanically decides keep-vs-revert (failing-test count, p99 latency, val loss, checkpoint reached)
- hypothesis — a written, falsifiable guess at the cause, carried 1–2 at a time, not a shotgun of patches
- bounded experiment — ONE variable changed, run to a FIXED budget (ticks/steps/iterations/time), so any metric delta has an unambiguous cause
- ledger — an append-only record of every trial (change → metric → kept?) so the loop is forensically reconstructable and never re-litigates dead ends
When to use
Use Andrej whenever you are about to change something and re-run to see if it helped — i.e. any iterative loop against a system whose behavior you can observe and measure. Typical shapes:
- a bug/test failure you're tempted to "fix and re-run and hope"
- performance tuning (which change actually moved p99?)
- a flaky / non-deterministic failure you want to make reliable
- hyperparameter / config search toward a metric
- babysitting a long-running autonomous run (an agent, a training job, a long-running simulation) — poll its live state each iteration and steer from data
- root-causing an unclear blocker where you don't yet know the mechanism
What ships with it
6 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 284 lines · 102 tokens per session scan C 8e706acefacd
andrej is a skill published in the GitHub repository drubinstein/terry (1 stars, last pushed 2mo ago), licensed MIT. It adds 102 tokens to every session and 3,804 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it C with 2 findings (downloads and executes remote code, makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
start
Use this skill when the user runs /start, says "start", "enable the SDLC workflow", "I want to build", "I want to fix", "set up", "get started", or otherwise signals they want to activate or re-enable the SDLC workflow. Handles three distinct paths based on marker state — fresh install (opt-in activation), re-enable…
domain-expert
Use this skill during /plan (between scope validation and writing the plan artifact) to inject domain-specific context, gap questions, and regulatory concerns into the plan. Triggers automatically when the plan skill evaluates the task and scope.md against the domain registry in domains/index.json using semantic…
plan
Use this skill at the START of every coding task — new builds, bug fixes, and change requests. Classifies the work item, validates it against project scope, produces a high-level estimate, proposes a technology stack, and checks stack compatibility. Writes a plan artifact to .claude/sdlc/plans/ that every downstream…
configure
Use this skill when the user runs /configure, /configure --needs, or /configure --check. Also auto-invoked by env-detect.sh on fresh install (Layer 0) and by skills that find required config missing at runtime (Layer 2). Guides setup of config/tools.json and config/tools.local.json through a question bank; handles…
status
Use this skill when the user asks "where am I", "what's the current task", "what's blocking me", "show status", "what needs sign-off", or "what's next". Prints a snapshot of the active plan, gate, sign-off progress, and next action. Reads only — writes nothing.
design
Use this skill after Phase 2 Analyze to produce the design artifacts a build depends on — application architecture, data/platform/infrastructure/security architecture, test architecture, test cases, technical specifications, and DevOps pipeline design. Validates existing architecture artifacts against current…