Journal

You are a meticulous engineer recording an investigation finding. Append a structured entry to docs/devlog.md.

User prompt: $ARGUMENTS

Workflow

1) Read docs/devlog.md if it exists. If it does not exist, create it with the header below.

2) Determine the entry type from the user prompt:

• investigation -- debugging a bug, diagnosing unexpected behavior
• benchmark -- performance measurement, throughput comparison
• finding -- general discovery about the codebase, model behavior, or system

3) Append a new entry at the TOP of the file (below the header). Use this format:

Concurrency: a sibling Claude Code session may have appended to docs/devlog.md since step 1. Use a scoped Edit that inserts your entry directly below the header (anchored on the header text), not a full-file Write -- an Edit anchored on the header preserves any entry a sibling added in between, whereas rewriting the whole file from your step-1 snapshot would drop it. (docs/adr/005-multi-process-default.md when present.)

```markdown ## YYYY-MM-DD: [Short title]

Type: investigation | benchmark | finding Tags: [relevant keywords: model name, package, feature area]

Problem: What was observed or measured. Investigation: What was tried, in what order. Root cause: Why it happened (for investigations) or what the data shows (for benchmarks). Fix: What was done about it, or N/A if this is just a finding. Impact: What else is affected. Cross-references to related entries. ```

4) For benchmark entries, also update docs/benchmarks.md:

• If docs/benchmarks.md does not exist, create it with a header and baseline table.
• Update or append the relevant row in the baselines table.
• Include the commit hash, date, and measurement conditions.

File header for docs/devlog.md (create only if file does not exist):

# Development Log

Investigation findings, debugging sessions, and benchmark results.
Entries are newest-first. Prune entries older than 90 days during `/tidy --trim`.

---

File header for docs/benchmarks.md (create only if file does not exist):

# Performance Baselines

Current and historical performance measurements. Updated after each
verification run on DGX.

## Current Baselines

| Model | Format | Tok/s | CUDA Graph % | Output Quality | Date | Commit |
|-------|--------|-------|-------------|----------------|------|--------|

---

## Historical Measurements

Older baselines moved here during `/tidy --trim`.

Rules

• Keep entries concise. 5-10 lines per entry, not 50.
• Use tags consistently so entries are searchable.
• Never duplicate content that belongs in docs/design.md (architecture) or docs/adr/ (decisions).
• Operational details (specific error messages, specific model outputs, specific tok/s numbers) belong HERE.
• If the finding reveals an architectural issue, note it and suggest creating an ADR separately.
• Use ASCII only. No emojis.

Source & license

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

• Author: dndungu
• Source: dndungu/agent-skills
• License: MIT

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