A command that records how a user changed a draft, such as rewriting, removing, or adding text. It saves these differences as records for later review.

It helps track changes to generated writing and find possible writing-rule updates based on repeated edits.

It turns manual edits into structured evidence instead of leaving improvements scattered across conversations or files.

learn — Command by Nisus74

Install

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.

agentmods

npx agentmods add commands/nisus74/humanise/learn

Clone the repo

git clone --depth 1 https://github.com/Nisus74/humanise

Per session 0 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.

When invoked 760 The whole file, excluding the scripts and references it only reads on demand.

Security scan A 0 findings. Scan, not verified.

Origin original No closer match found in the catalogue.

Token cost

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	$0.00000	$0.00760
Opus 5	$0.00000	$0.00380
Sonnet 5	$0.00000	$0.00152
Haiku 4.5	$0.00000	$0.00076

Measured yesterday against content hash e812e84923c0, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

learn 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 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.

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.

skill/commands/learn.md · 44 lines

How it starts

The opening of the file, as written. The whole thing — 44 lines — stays where its author put it; the contents beside it link to each section on GitHub.

/humanise learn

Capture what the user changed in a draft the skill wrote, so the loop can learn from it. Every rewrite, cut, or addition is evidence; this command turns it into ledger records that /humanise improve mines for rule-change candidates.

Inputs

Two texts:

The skill's draft. From this session if it's still in context, or a file the user points at.
The shipped text. What the user actually sent or published: pasted into chat, or a file path.

Write each to a temp file if it isn't one already. Ask for the channel and audience tag if the session doesn't already know them; unknown is allowed but clusters worse.

Steps

Run the capture script:
```
python3 scripts/capture_edit.py --draft <draft> --final <final> --channel <channel> --audience <tag>
```
It diffs the texts sentence by sentence, classifies each changed span against the deterministic checks, and appends records to profile/learning/ledger.jsonl (the soul; gitignored, never ships). The script prints every record it wrote.
Classify the unexplained spans. Any record printed as voice (unclassified) is an edit no deterministic check explains. Read the draft span next to the final span and pick the mechanism from this vocabulary (a fixed menu is what keeps the ledger clusterable):
- too-even-rhythm: the user broke up same-length sentences or added a fragment.
- no-stance: the user inserted an opinion or a call the draft ducked.
- too-generic: the user swapped an abstraction for a number, a name, or a date.
- synonym-cycling: the user collapsed elegant variation back to one repeated word.
- opener-template: the user rewrote a stock opening or closing move.
- register-miss: the user shifted formality up or down (see references/tone-register.md).
- wrong-fact: the correction is about content rather than voice; usually not a rule candidate.
- other:<slug>: nothing fits; coin a short slug and say why in note.
Append one superseding record per classified span: same span_id, source: "memory", the chosen mechanism, and a one-line note. Append with the same JSON shape the script wrote; never edit existing lines. Mining takes the latest record per span, so the classification wins without rewriting history.

Read the full file on GitHub · 44 lines

Changes

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 · 44 lines · 0 tokens per session scan A e812e84923c0

Subscribe to this mod's changes

learn is a command published in the GitHub repository Nisus74/humanise (1 stars, last pushed 5d ago), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 760 tokens. 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.

Code review Writing and docs agent-skills ai-tools ai-writing claude-code codex cursor

learn

Getting it into your agent

What it costs to keep this loaded

Grade A, and why

Nothing flagged

How it starts

/humanise learn

Inputs

Steps

What this file has done since we first saw it

Other commands, from other repositories

test-e2e

export

newpkg

tray-referencia-api

remove-bg

perf-investigate