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 commands/zircote/human-voice/voice-driftgit clone --depth 1 https://github.com/zircote/human-voiceWrote 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/commands/zircote/human-voice/voice-drift)<a href="https://agentmods.dev/commands/zircote/human-voice/voice-drift"><img src="https://agentmods.dev/badge/commands/zircote/human-voice/voice-drift.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 | $0.00030 | $0.00496 |
| Opus 5 | $0.00015 | $0.00248 |
| Sonnet 5 | $0.00006 | $0.00099 |
| Haiku 4.5 | $0.00003 | $0.00050 |
Grade A, and why
voice-drift 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 5d 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 — 40 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Voice Drift Report
Generate a drift report comparing accumulated voice observations against the current profile.
Procedure
-
Recall observations: Search Atlatl memory for voice observations:
recall_memories(query="voice drift observed pattern writing", namespace="voice-observations", mode="hybrid") -
Load profile: Read
~/.human-voice/profile.jsonfor current dimension scores. Extract the score for each dimension from thedimensionsobject. Do not use hardcoded values; every user's profile is different. -
Compare: For each observation, compare against the dimension scores loaded from the profile in step 2. Cover all gold standard dimensions (formality, emotional_tone, personality, complexity, audience_awareness, authority, narrativity, humor) and all gap dimensions that have scores. Also compare mechanics observations (contractions, Oxford comma, punctuation style) against the
mechanicssection of the profile. -
Report in three sections:
Confirmed patterns: Observations that align with the profile. The user writes as the profile predicts.
Drift detected: Observations that consistently diverge from the profile. Include the dimension name, the profile score, the observed tendency, the direction of drift (warmer, more formal, shorter sentences, etc.), and the number of observations supporting the pattern.
New patterns: Voice characteristics observed that are not captured in the current profile at all.
-
Recommend:
- If no drift: "Profile remains accurate. No action needed."
- If moderate drift (1-2 dimensions): "Consider updating the profile for [dimensions]. A targeted re-interview of [modules] would refine the scores."
- If significant drift (3+ dimensions): "The profile may no longer represent your current voice. Consider a full re-interview with /voice-interview."
-
If no observations exist: "No voice observations have been recorded yet. The observer protocol accumulates observations silently as you write during Claude Code sessions. Continue working and check back after you have authored several pieces of content."
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.
- 5d ago First seen · 40 lines · 30 tokens per session scan A 95e7865b0142
voice-drift is a command published in the GitHub repository zircote/human-voice (10 stars, last pushed 1mo ago), licensed MIT. It adds 30 tokens to every session and 496 once invoked, about $0.0002 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 commands, from other repositories
graph
Extract prerequisite concept graph from converted course materials and write course-index/concept-graph.md.
reindex
Idempotent reindex of an existing course — rewrite retired coverage.md tier markers to canonical vocabulary and materialize errors/log.md nature/phase into the on-disk schema (header keys == data keys), in place and atomically, WITHOUT running analyze.
alt
Import an Exam Radar (OPTIMETA Alt plugin) export and fold its lecture-emphasis exam signal into the course index — radar.md, a lecture-emphasis column on coverage.md, and a gold-zone weakmap.
blind
Strategy-level blind drill on a known HW or example problem. User describes approach in prose (no math typing); Claude verifies against solution then saves clean reference to derivations/.
doctor
Diagnose the paideia install + course workspace (Python, poppler, tesseract, Ollama/Qwen3-VL, course folders, .course-meta, writable paths, statusline wiring) and optionally auto-repair the permission-free issues. Run when a command won't work or right after cloning.
quiz
Generate N practice problems on a topic. Saves problem MD + hidden answer MD. User solves on paper, uploads answer PDF, then runs /grade. First arg weakmap → auto-load latest weakmap report and target its top weaknesses.