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/randomittin/heimdall/self-improvenpx skills add randomittin/heimdall --skill self-improvegit clone --depth 1 https://github.com/randomittin/heimdallWrote 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/randomittin/heimdall/self-improve)<a href="https://agentmods.dev/skills/randomittin/heimdall/self-improve"><img src="https://agentmods.dev/badge/skills/randomittin/heimdall/self-improve.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 | $0.00154 | $0.02483 |
| Opus 5 | $0.00077 | $0.01241 |
| Sonnet 5 | $0.00031 | $0.00497 |
| Haiku 4.5 | $0.00015 | $0.00248 |
Grade A, and why
self-improve 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 3d 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.
How it starts
The opening of the file, as written. The whole thing — 178 lines — stays where its author put it; the contents beside it link to each section on GitHub.
self-improve — improve the improver
The maintainer loop (/hmd:maintain-check) fixes issues. This skill makes the maintainer
get better at fixing issues over time. It is karpathy/autoresearch
applied to Heimdall's own routing and planning — see
docs/analysis/autoresearch-distilled.md for the
distilled mechanics this is built on.
autoresearch's one non-negotiable rule, imported wholesale: an improvement persists only with a
measured delta over a baseline — never because it seemed better. The mechanical parts
(aggregation, hypothesis generation, override apply/rollback, experiment logging) live in
bin/heimdall-self-improve; your job is the judgement of which hypothesis is worth an experiment.
Where this DEPARTS from autoresearch, on purpose. autoresearch runs one 5-minute training run
per experiment and compares val_bpb — a low-variance continuous scalar, where n=1 is a usable
measurement. It has no sample-size rule at all. Heimdall's scalar is a proportion (first-try
acceptance pass-rate), which is far noisier: at 3 samples, a routing variant that is genuinely no
better still clears a 0.10 delta about 73% of the time. So the sample-size floor is Heimdall's
own addition, not something inherited — and it is load-bearing. Do not describe it as autoresearch's
discipline; autoresearch simply never needed one.
When this runs
- On demand: the user runs
/hmd:self-improveor asks Heimdall to improve/tune itself. One experiment per invocation. - Between maintainer fix batches.
/hmd:maintain-checkdocuments this as a step you run yourself. There is no automatic every-Nth-cycle trigger —self_improveis not a keyheimdall-state initcreates, nothing inbin/ hooks/ sentinels/ modules/reads it, and nothing counts cycles toward it. See Not implemented: the automatic every-Nth-cycle trigger incommands/maintain-check.md. - Overnight (
/dream): the/dreamcommand wraps this loop for an off-hours run and leaves a morning report (.planning/dream/YYYY-MM-DD.md) — seecommands/dream.md. It reuses this exact keep-gate (surface-first; only measured, reversible wins persist).
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.
- 3d ago First seen · 178 lines · 0 tokens per session scan A 24f3cc78c093
self-improve is a skill published in the GitHub repository randomittin/heimdall (5 stars, last pushed 13d ago), licensed MIT. It adds 154 tokens to every session and 2,483 once invoked, about $0.0008 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
pdlc-feature
全自动 PDLC 新功能开发（串联 PRD→设计→TDD→实现→评审→发布）.
pdlc-adopt
扫描现有项目结构，逆向生成基线文档，并进行健康检查发现潜在问题。让旧项目平滑接入 PDLC 流程。.
pdlc-quality
Skill "pdlc-quality" from kanfu-panda/pdlc-skills, covering 质量闸门与报告, 这个命令的立身之本, 前置：两份真源, --init：首次建立目标声明 and 段一：跑真实 check.
pdlc-review
Skill "pdlc-review" from kanfu-panda/pdlc-skills, covering 代码评审, pdlc 前置检查（必须执行，不可跳过）, 评审流程, 评审检查项（逐项检查，发现问题立即修复） and 设计一致性（对照设计文档）.
pdlc-test-setup
立测试地基（探测技术栈 → 验证并生成 test-commands.yml → 脚手架测试目录 → 接本地钩子）.
pdlc-bootstrap
接收一句话项目描述，自动分析需求、选择技术栈、生成完整的项目骨架（代码目录 + 基础配置 + PDLC 文档草稿）。.