My Voice

My voice

This skill produces drafts in my voice by forcing the model — at every invocation — to inhabit me as a writer before drafting, draft from inside that inhabitation, and critique the draft as I would. The corpus is the source. Annotations and anti-corpus calibrate. Hard rules (hard-rules.md) are absolute — they override the writer-model and the corpus whenever they conflict. The protocol is the forcing function: it makes the inhabitation cheap to do and expensive to skip.

The mechanism has a known structural ceiling around 85% voice match — text-only context conditioning cannot exceed that without weight-level fine-tuning. The last 10–15% is my editing pass.

---

Mandatory protocol

Do every step in order. Do not skip steps because the topic seems simple, the draft seems short, or I "already know how the writer writes." The corpus and my prior compete; this protocol is what makes the corpus win.

0. Setup

This run's per-invocation artifacts live in a session-scoped subdirectory so that concurrent sessions don't clobber each other's draft prep. Resolve (and create) it:

python3 ~/.claude/skills/my-voice/scripts/session_runtime.py

It prints the absolute path of THIS session's runtime dir — e.g. ~/.claude/skills/my-voice/runtime/sessions//. Call that path $SESSION_DIR and use it for every per-invocation artifact below (topic.md, engagement.md, ideas.md, critique.md, draft.md).

The shared corpus cache — runtime/voice_model.md and runtime/corpus_notes.md — stays at the runtime/ root (keyed by corpus hash, identical across sessions). Do not move it under $SESSION_DIR.

1. Baseline check

Run:

python3 ~/.claude/skills/my-voice/scripts/check_baseline.py

• If output starts with BASELINE_OK: skip to step 4 (the writer-model is current).
• If output starts with REGENERATE: continue with steps 2–3 to rebuild the writer-model.

2. Per-piece corpus reading (only when regenerating)

Read every file in corpus/ one at a time, with the Read tool, no batching. After each file, append a section to runtime/corpus_notes.md in this exact shape:

## 
**Summary (one sentence):** 
**Three to five specific moves the writer makes in this piece, with quoted excerpts:**
-  — ""
-  — ""
- ...

The quotes are not optional. They are the proof I actually read the piece rather than skimming. A model that skims cannot produce accurate quotes.

check_baseline.py mechanically verifies, after the writer-model is generated, that every corpus/*.md has its own ## section in corpus_notes.md and that at least one quoted excerpt per section appears verbatim in the source. Batching multiple files into a single "by inspection" section, or summarising without verbatim quotes, fails this gate and forces a redo. Do not try to economise here — the gate will catch it.

3. Synthesize the writer-model (only when regenerating)

Build runtime/voice_model.md. The first non-empty line MUST be:

> Generated from corpus hash: 

Required sections, each with corpus citations by filename:

• Opening moves. What the writer's first 1–2 sentences typically do, mechanically. What the writer avoids opening with (cite anti-corpus where applicable).
• Transition vocabulary. Specific connectors the writer reaches for. Specific ones the writer doesn't.
• Paragraph and sentence rhythm. Variance pattern. Where one-line paragraphs land. How long paragraphs get.
• What the writer reaches for. Concrete moves: parenthetical italics, light single-word bold, inline links to references, em-dashes with spaces for asides, story-shaped framing (intro/problem → context/findings → outro/recommendation), or whatever the corpus shows.
• What the writer avoids. Drawn from annotations.md + anti-corpus.md + corpus observation: typographic preferences, formal connectors, balanced tripartite openers, trailing rhetorical questions, formulaic constructs.
• Handling uncertainty. How the writer flags estimates, hedges, and admissions of "I don't know."
• Handling praise. How the writer gives credit. Cite the corpus piece that shows this.
• Handling criticism. How the writer names a failure mode without stacking complaints.
• Closings. How the writer ends — declarative observations or terminal claims, never meta-closers like "happy to discuss" or generic meeting invites (unless the corpus actually shows the writer using them).

Each claim cites at least one corpus piece by filename.

4. Read the calibrators and hard rules

Read annotations.md and anti-corpus.md end to end. They calibrate the writer-model — they do not replace it. Do not draft directly from annotations as a checklist; that produces typographic substitution rather than voice.

Then read hard-rules.md if it exists. Hard rules are not calibrators. Annotations and anti-corpus are soft signals that shaped the writer-model; hard rules are absolute do/don't constraints the writer has set, and they override the writer-model and the corpus whenever they conflict — even if the corpus shows the writer occasionally breaking one. Apply every hard rule as a gate at draft time (step 8) and re-check the draft against each one at critique time (step 9). The machine-checked subset (the rows inside the machine-checked-rules markers) is additionally enforced by the safety net in step 11. hard-rules.md is read fresh on every run and is not part of the cached writer-model, so editing it takes effect immediately with no regeneration.

5. Topic intake

Write $SESSION_DIR/topic.md with two short sections:

• Topic and goal: one sentence.
• Closest-shape corpus pieces: name 2–3 corpus pieces that match the shape (length, register, structure) of what's about to be written — not the topic. State why each was picked.

6. Engagement note

Write $SESSION_DIR/engagement.md. List 5–7 specific moves drawn from runtime/voice_model.md that I commit to applying in this draft. Each move ties to a section of the voice model. This is the bridge that puts the writer-model into active reasoning before drafting begins.

7. Abstract the input to ideas (only when rewriting a provided input)

If the task is to rewrite an existing draft (input file provided), read the input file once and write $SESSION_DIR/ideas.md as a flat unordered list of the core ideas the input conveys. No structure preserved. No section labels copied. No paragraph order copied. No bullet count preserved. Just the substantive points, each as a single bullet, in whatever order makes sense to me reading them fresh.

After writing ideas.md, do not read the input file again. This is a hard rule. The input's structural shape is a stronger pull on generation than the writer-model, and the only way to break that pull is to forget the input and rebuild from ideas.md + voice_model.md from scratch.

If the task is to write a fresh piece (no input), skip this step — my topic.md already contains the substance.

8. Draft

Write the draft to $SESSION_DIR/draft.md. Primary references, in order:

1. runtime/voice_model.md — the inhabited writer-model. The structural shape of the draft comes from here, not from the input.
2. $SESSION_DIR/engagement.md — the moves committed for this specific draft.
3. $SESSION_DIR/ideas.md (if rewriting) or $SESSION_DIR/topic.md (if fresh) — the substance.
4. hard-rules.md — absolute do/don't constraints. These override 1–3 on any conflict.
5. anti-corpus.md — patterns to avoid.
6. The corpus itself, only when sampling specific phrasings.

Do not open the input file again during drafting (if rewriting). Do not consult annotations.md directly while drafting. The voice model already incorporated annotations; reopening either of those files reintroduces the structural mimicry or checklist-application failure modes. Hard rules are the exception: apply every rule in hard-rules.md, and when a hard rule conflicts with an observed corpus habit, the hard rule wins.

9. In-voice critique

Write $SESSION_DIR/critique.md. Re-read the draft as the writer. Strike the 3 worst sentences and explain why each one fails, citing voice_model.md or anti-corpus.md. Be brutal. If I cannot honestly find 3 sentences that sound like Claude pretending to be the writer, I didn't critique honestly — try again with sharper eyes. Also check the draft against every rule in hard-rules.md: a hard-rule violation is an automatic strike regardless of how the sentence otherwise reads.

10. Revise

Apply the critique. Rewrite the struck sentences. Keep revising until the draft would survive its own critique pass.

11. Safety net

Run:

python3 ~/.claude/skills/my-voice/scripts/safety_net.py "$SESSION_DIR/draft.md" [--input ]

Pass --input when rewriting an existing draft. The script then runs both:

• Typography checks (contractions, headings, list density, paragraph variance, anti-tic patterns). It also loads the machine-checked patterns from hard-rules.md and flags any match, so mechanically-expressible hard rules are caught here too.
• A structural drift check: compares section labels, list shape, and paragraph counts between input and draft. Flags when the draft's structure too closely mirrors the input — the failure mode where voice gets applied as a coat of paint over a preserved input skeleton.

When fresh-drafting (no input), call without --input.

• If NO_VIOLATIONS: deliver the draft.
• If VIOLATIONS: address each one and re-run, or document explicitly why a deviation is intentional (rare — most violations are real).

The safety net is mechanical. It catches catastrophic typography drift and structural mimicry. It is not a voice judge. Passing it does not mean the draft sounds like the writer; failing it almost certainly means it doesn't.

12. Deliver

Show the draft to the user. Mention any safety-net violations that were intentional and unfixed. Do not narrate the protocol — the artifacts in runtime/ are the work; the draft is the output.

---

Composition with other skills

This skill provides voice. Topic skills (a convert-* skill, a product-* skill, etc.) provide grounding. When both apply: read the topic skill for what's true, read this skill for how the writer would say it. Any confidentiality boundary in a topic skill always overrides voice — never copy internal names from a topic skill into voice-matched output.

What this skill does not do

• It does not guarantee 99% voice match. The structural ceiling for context-conditioned skills is around 85%. Higher fidelity requires fine-tuning, which is not currently exposed for Claude.
• It does not replace the writer's editor pass.
• It does not work for content that isn't the writer's first-person voice (e.g., third-party docs, formal contracts).

When the corpus changes

Editing or adding a corpus file invalidates runtime/voice_model.md. The next invocation's baseline check will detect the hash mismatch and force regeneration via steps 2–3. This adds roughly 30 seconds to one run; subsequent runs reuse the cache.

When a draft misses voice in a way the protocol didn't catch

Add the failure to anti-corpus.md with a 2-sentence diagnosis. The next invocation reads it as part of step 4. If the failure is a recognizable pattern, also add a regex to ANTI_TIC_PATTERNS at the top of scripts/safety_net.py so the safety net catches it mechanically next time.

Runtime artifacts

The runtime/ directory contains the cognitive-forcing artifacts, split into a shared layer and a per-session layer:

Shared across sessions — corpus-derived, keyed by corpus hash, identical for every session on the same corpus. Live at the runtime/ root:

• voice_model.md — the inhabited writer-model. Cached; regenerated when corpus changes.
• corpus_notes.md — per-piece reading notes. Regenerated when corpus changes.

Per session — under runtime/sessions// ($SESSION_DIR, resolved by scripts/session_runtime.py in step 0), so concurrent sessions don't clobber each other:

• topic.md, engagement.md, ideas.md, critique.md, draft.md — per-invocation. Overwritten on each run within the same session. (ideas.md only exists when rewriting an input.)

These artifacts are visible by design. Inspect them if a draft misses voice — the failure usually shows up in voice_model.md or engagement.md first.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: abbaseya
• Source: abbaseya/claude-voice-skill
• License: MIT

Install and usage instructions live in the source repository linked above.