Free capture kit · schema version 1

Portable structured experiment records

Generate one local Markdown file per experiment with Claude Code, Codex, Cursor, Gemini CLI, Copilot, Kimi, a notebook, or no agent at all. Then upload it to a deterministic preview, correct every mapped field, and save only a draft.

MIT-licensed generic package. It contains no customer-contributed content. Obtain explicit written permission before redistributing any customer-provided skill, hook, text, or adaptation.

Version 1 file shape
---
schema_version: 1
date: 2026-08-03
project: Adaptive sensor calibration
topic: Thermal drift recovery under rapid cycling
author: Jordan Lee
source_tool: claude-code
source_generated_at: 2026-08-03T16:42:00-07:00
total_session_hours: 3.0
sred_hours: 2.5
---

## Technological uncertainty
## Hypothesis
## Work performed and experiments
## Observations and results
## Conclusion and technological learning
## Remaining uncertainty or next experiment
## Evidence references

Deterministic first

Explicit front matter and headings are parsed without asking an AI to interpret or follow the uploaded text.

Human-reviewed drafts

Every field can be edited or remapped. Imports never become approved experiments automatically.

Provenance retained

The exact original, content hash, upload time, claimed generation time, producer, duplicate signal, and audit history stay attached.

What is in the download

Capture files

  • Tool-neutral experiment-logging skill
  • Optional local end-of-session reminder hook
  • Blank template and complete failed-result sample
  • Schema contract and compatibility checklist

Operational guardrails

  • Never transmits automatically
  • Hours are optional and never guessed
  • Underlying commits, tests, issues, notebooks, and datasets remain the evidence
  • Change log protects version compatibility
Human review is mandatory. An AI-generated session summary is not complete proof, tax advice, or an eligibility conclusion. Source generation time is a claim from the file; SR&ED Copilot records the later upload time separately. Failed and inconclusive learning is valid technical context and should not be rewritten as success.