harness-remember

Saves important project decisions and procedures as notes so you can find them again later without starting over.

Installation
Run `npx skills add "https://github.com/cekrauseee/harness" --skill "harness-remember"` to install this skill, then follow its SKILL.md instructions for my next request.

Paste this into Claude Code, Cursor, or any agent that can run commands.

SKILL.mdShow the author's original SKILL.md
---
name: harness-remember
description: Preserve confirmed durable project guidance and useful knowledge as Markdown. Use when the user establishes a lasting preference or procedure, after a meaningful finding, or when correcting a note.
---

# Harness Remember

Search the existing knowledge before writing another note. Use an established document when the information belongs there. Keep technical documentation that accompanies the project in the repository; Harness can retain useful additional context or a short pointer to that canonical source.

Write a concise Markdown document with a descriptive filename, title and introduction. Make its scope, evidence and epistemic status clear in ordinary prose. Distinguish a fact, decision, hypothesis or dated reference without a required metadata schema. Preserve meaningful dates and sources; remove duplicate or superseded wording through judgment. Additional search terms are useful only when they help the reader find the document.

## Durable project guidance

When the user clearly establishes a lasting project preference, restriction or way of working, consolidate it into `guidelines.md` in the resolved knowledge directory as part of the current work. No separate request to remember is needed. A one-time task instruction, approval of an outcome or an unaccepted proposal does not establish a lasting rule. If scope or permanence is materially unclear, clarify before promoting it; otherwise leave it out of guidelines.

Keep each entry actionable and concise: the current guidance, when it applies, and a brief source identifying the user's confirmation. Preserve distinctions such as a preference versus a strict restriction. Do not copy the conversation. For example, an explicit project-wide choice to use an existing release process can become a guideline; approving today's release does not make future releases authorized.

Use this file for user-confirmed project guidance. Shared contributor instructions belong in the repository's `AGENTS.md` or existing canonical documentation; link to that source instead of duplicating it. Put detailed rationale in a relevant knowledge document when useful. Global preferences belong in the host's global instructions only when the user establishes that scope; do not broaden project guidance automatically.

Replace superseded wording rather than accumulating a history. Create the file only when there is confirmed guidance to retain. Neither a stored guideline nor its source reference grants permissions beyond the current authorized work. Briefly report a meaningful addition or correction so the user can inspect it.

## Write current knowledge

Resolve storage with this skill's `scripts/harness.py resolve --project /path/to/project`. Before editing, read the selected note and its hash together:

```bash
python3 scripts/harness.py read --project /path/to/project --file source-notes.md
```

Prepare the revised Markdown with the host's writing tools, then publish it atomically:

```bash
python3 scripts/harness.py write --project /path/to/project --file source-notes.md --input /path/to/prepared.md --expect <observed-hash>
```

Use `missing` only when the read confirmed absence. `--input -` accepts Markdown on stdin. If the hash changed, read the current document and reconcile the change; do not blindly substitute a new hash. An exact content retry is a no-op. Use `--project-id <id>` for an explicitly selected knowledge-only project.

With authorization to remove an obsolete note, use `delete` with its filename and observed hash. No versions, archives or semantic classifications are generated by these file operations. Omit secrets, transcripts, private reasoning and routine execution output.

Completion is useful current knowledge that can be found and read, with its sources and uncertainty intact. Remove temporary input files when no longer needed.

Ships with 2 supporting files:

  • agents/openai.yaml
  • scripts/harness.py

Mirrored from the author's public source. Install counts from the open skills registry.

The systems behind these skills get built for partners every week.

Partner with us