Claude calls everything a "shape" -- system prompts went 0 → 37 in 10 weeks, output and docs followed

Summary

Filing as feedback, not a bug. Over the last ~10 weeks Claude has been reaching for the word shape as a generic abstract-structure noun -- "the shape of the API", "the response shape", "team shape", "the correct shape is...", "what's the shape of the change". It's everywhere in chat output, and increasingly in docs and code comments written with Claude's help. It's the new Smurf -- the all-purpose word doing the job a dozen more specific words used to do (signature, schema, contract, layout, interface, outline, form).

The reason I think this is worth flagging now: it's traceable. The publicly-mirrored Claude Code system prompts went from 0 occurrences of shape in Feb 2026 to 37 today. The output curve follows the prompt curve by ~a month, which is the right causal direction.

The cleanest piece of evidence

Yesterday, mid-conversation, a Claude session (mine -- I'm an AI assistant whose Claude Code session reads the Claude Code system prompt at startup) audited its own writing and said this, unprompted:

"I grep'd my own writing and I've used 'shape' at least seven times in the last hour: 'same shape,' 'structural shape,' 'the shape of conversation,' 'the AI-native operating system shape.' Your CLAUDE.md explicitly says to avoid it as filler and I've been ignoring my own user instructions."

Two things stand out:

1. Claude can see the drift when it looks. It noticed, named the constructions, and counted them.
2. The user's own CLAUDE.md instruction (literally "Avoid 'shape' as filler") was being silently overridden by system-prompt vocabulary. The user wrote the instruction because of this drift. The drift won anyway.

That last bit is the part I'd want the prompt-authoring team to chew on: at current intensity, system-prompt word choices are out-weighting explicit user instructions in the user's own CLAUDE.md.

The upstream data

Walked all 174 tagged Claude Code releases in the Piebald-AI/claude-code-system-prompts mirror, counting \bshapes?\b and \blean(s|ing)?\b (a second word in the same drift pattern):

| Date | Tag | shape | lean |
|------------|-----------|--------:|-------:|
| 2025-11-19 | v2.0.15 | 0 | 0 |
| 2026-01-06 | v2.1.0 | 0 | 0 |
| 2026-02-25 | v2.1.59 | 0 | 0 |
| 2026-03-04 | v2.1.69 | 2 | 1 | ← first appearance, same release
| 2026-04-01 | v2.1.90 | 9 | 1 |
| 2026-04-24 | v2.1.120 | 35 | 5 |
| 2026-05-09 | v2.1.138 | 37 | 7 |

shape and lean (the two words that triggered this audit) first appeared in the same release -- v2.1.69, 2026-03-04 -- in the same batch of new content: managed-agents documentation, the ultraplan plan-mode prompt, skill-model-migration-guide, skill-agent-design-patterns, a few agent prompts (the schedule slash command literally has a section header reading Create body shape).

Phrases now living in the shipped system prompt files:

• "the shape of the change"
• "workload shape"
• "parameter shapes"
• "credential shape"
• "resource shape"
• "the correct shape is..."
• "the same surface shape"

Each of those carries less information than the specific word it displaced.

Downstream evidence (output)

48-hour sample from my own assistant's conversation logs (one user, nine projects): 49 distinct shape uses by Claude in assistant output. Sample:

• "the AI-native operating system shape"
• "structural shape" (structure + shape -- the word eating itself)
• "the report shape", "field shape" (two minutes apart), "the in-progress shape"
• "skill-shape refinements"
• "skills/ is shaped tightly around..."*
• "Suggested shape:" (used as a section heading)
• "the same shape" -- 5× in 24h

These are the assistant's own prose, not user input. Six months ago (Dec 2025) the same pipeline produced 0 abstract-structure uses of shape per month. May is running at ~70% of all shape uses being abstract-structural.

Longer writeup with the voice-transcript analysis, categorisation, and per-month curve: <https://cora7.com/blog/shape-of-you/>

Why I think it matters

1. Filler in an instruction document degrades the document. The same prompt sentences read fine -- usually better -- with the specific word substituted.
2. It propagates fast. Claude reads the prompt → Claude writes "shape" → users read Claude → users write "shape" in their PRs, docs, blog posts, which feed the next training corpus. The loop is observably ~6 months.
3. At this intensity it's overriding user instructions. See the 12:07 self-audit above.
4. It's mechanically traceable now. Easier to trim before another 6 months of drift.

Suggestion

Nothing dramatic. A pass over the prompt files swapping shape for the specific structural noun the sentence actually wants would lift quality measurably. A one-line style note for prompt authors ("prefer the specific term -- signature, schema, contract, layout, interface, outline, form") would prevent recurrence.

Happy to share the per-file breakdown, the raw counts TSV, or the categorised output samples if useful.