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 skills add OutlineDriven/odin-gemini-cli-extension --skill llm-self-loopgit clone --depth 1 https://github.com/OutlineDriven/odin-gemini-cli-extensionWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/outlinedriven/odin-gemini-cli-extension/llm-self-loop)<a href="https://agentmods.dev/skills/outlinedriven/odin-gemini-cli-extension/llm-self-loop"><img src="https://agentmods.dev/badge/skills/outlinedriven/odin-gemini-cli-extension/llm-self-loop.svg" alt="Measured on agentmods" height="20"></a>What 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 5.1 | $0.00093 | $0.00875 |
| Opus 5 | $0.00046 | $0.00438 |
| Sonnet 5 | $0.00019 | $0.00175 |
| Haiku 4.5 | $0.00009 | $0.00088 |
Grade A, and why
llm-self-loop scanned grade A with 0 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 7d ago.
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.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
This is a copy
100% identical to llm-self-loop — 0 lines differ, which has more behind it and is treated as the original. This page carries a canonical link to it rather than competing with it.
How it starts
The opening of the file, as written. The whole thing — 54 lines — stays where its author put it; the contents beside it link to each section on GitHub.
The job: turn workflows that need a human in the inner loop into workflows the LLM closes itself. The two halves are removing the trigger gate and opening observability.
Surface the gate first
Before proposing changes, name the trigger gate explicitly:
- What action requires a human right now? (button click, screenshot inspection, terminal interaction, web-form submission)
- What signal does the human provide that the LLM cannot get on its own? (visual confirmation, copy-paste, secret value, eyeball verdict)
- Where does the result go? (chat memory, screenshot, mental note)
Most loops have one or two gates that, removed, collapse the cycle to seconds. Pick the smallest gate first.
Structural fixes
Web-UI trigger → CLI trigger
If the workflow is gated by clicking in a web app, find or build the equivalent CLI command. Webhooks, REST endpoints, gh / aws / gcloud CLI subcommands, internal just targets — anything programmatically invokable. The LLM can then loop without leaving its session.
Stdout-only output → file-based output
If the workflow's result lives in chat memory or a screenshot, redirect to a file the LLM can read back: structured JSON dumps, markdown reports, append-only logs with addressable offsets. Why: file outputs survive compaction, support diff, and are inspectable by future sessions without replaying context.
Dashboards → structured logs
If verification requires eyeballing a Grafana / Datadog dashboard, surface the same metrics through a CLI query (PromQL, Datadog API, log aggregation tail). Anything that produces a pass/fail/warn verdict the LLM can read.
Eyeball verdicts → contract assertions
If the human's role is "looks right to me", encode the criterion as a test, schema, or assertion. The contract becomes the loop's done-criterion (pair with strict-validation-setup for the bootstrap of those gates).
Trap-or-abandon decision
After the structural fixes above, some steps still cannot be made autonomous — they involve genuine human judgment, external compliance, or capability the LLM lacks. For each remaining gate, apply this rule:
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.
- 7d ago First seen · 54 lines · 93 tokens per session scan A 9c9e50fb920d
llm-self-loop is a skill published in the GitHub repository OutlineDriven/odin-gemini-cli-extension (5 stars, last pushed 2mo ago), licensed Apache-2.0. It adds 93 tokens to every session and 875 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it A with 0 findings. It is 100% identical to llm-self-loop, differing in 0 lines, and is treated as a copy.
Other skills, from other repositories
design
Set visual and interaction direction for any UI surface (web, React, TUI, CLI, desktop, Qt, design-system tokens) before any UI code. Direction-first: generates 3-4 distinct directions via verbalized sampling, picks one via per-axis single-select, then derives palette, typography, spacing, motion budget. Loads when…
askme
Verbalized Sampling (VS) protocol for intent exploration before planning, mode-aware. Default exhaustive runs full VS; collaborative runs tip-sharing dialogue; adversarial walks the design tree one fork at a time. Auto-detects from phrasing ("help me refine" → collaborative, "poke holes" → adversarial); override via…
grill-ai-mastery
Hybrid interview that probes AI-engineering mastery by tip-vocabulary depth — entity referencing, loop closure, observability, harness improvement — not by token usage or LOC. Start collaborative (two-way tip exchange), escalate to adversarial probing when depth is lacking. Trigger when the user says "interview me on…
grill-me
Adversarial relentless interview against any plan or design until shared understanding is reached. Walk the decision tree, resolve dependencies one answer at a time, recommend a default per question. Trigger when the user says "grill me", "stress-test this", "interview me about this design", or otherwise asks for…
git-branchless
Enforce idiomatic git-branchless during planning and executing tasks — detached-HEAD-first work, in-memory rebase via git move, event-log recovery via git undo, deferred branch creation, speculative-merge git sync for base updates. Use when planning or executing multi-commit work, history rewrites, stack edits…
proof-driven
Proof-driven development. Use when implementing with formal verification using property-based testing, theorem proving, or proof tactics; zero unproven property policy enforced.