Kaysser Kayyali
e1a0e836ac
The executable mirror in prompts/system-prompt.md drifted from 07-reasoning-harness.md after the recall reframe. The skill mandates the two stay in sync (drift caught by test_system_prompt_consistency in slice 7). Mirror now matches: recall is the fallback because un-typed/un-cited/un-time-bounded, not low-precision. Co-Authored-By: Claude <noreply@anthropic.com>
prompts/
Status: 📋 planned. The prompt registry.
Convention
- One file per prompt, markdown with a YAML frontmatter header.
- Filename is the prompt's stable id
(e.g.,
system-prompt.md,extraction-prompt.md). - Frontmatter holds:
id,version,used_by,last_eval,eval_pass_rate. - Prompts are versioned in git. Bumps follow semver:
- patch: typo fixes, comment-only changes.
- minor: wording changes that don't shift behavior.
- major: any change that might shift behavior.
- Major and minor bumps require the harness to re-run
(
docs/18-eval-policy.md).
File template
---
id: <stable-id>
version: <semver>
used_by: [<slice or system that loads this prompt>]
last_eval: <date>
eval_pass_rate: <0.0–1.0, or "n/a">
---
<the prompt text, verbatim>
Current prompts
| File | Used by | Status |
|---|---|---|
system-prompt.md |
The reasoning harness | mirror of docs/07-reasoning-harness.md |
extraction-prompt.md |
cognee.cognify override |
stub — slice 3 writes the real prompt |
Why a registry, not embedded prompts
- Prompts are load-bearing. They need a home separate from the design docs that describe them.
- A/B testing two prompt variants is
git checkout v1 vs v2, not a hunt through markdown. - Eval pass rates belong next to the prompt, not in a spreadsheet.
Cross-references
docs/07-reasoning-harness.md— the system prompt source-of-truth (this directory is the executable mirror)docs/18-eval-policy.md— when prompts get re-evaluateddocs/cognee-integration.md— where the extraction prompt is loaded