Step 1 / 5 · 15 min
Know whether context changed the result.Freeze the baseline
- Use a representative input.
- Save prompt, output, defects, score, and review time.
Build · Core skill lab
Before / After Context Pack
Improve an AI task by changing the information architecture, then prove the pack reduces defects or review time.
- Decision
- Which context materially improves this recurring task—and which context should be removed?
- Timebox
- 75 minutes + two comparable runs
- Output
- Context experiment report
Bring to the bench
- One recurring AI task
- A representative input
- A baseline output you can score
Your work saves in this browser.
Field tools
Use the instrument, not a blank page.
Copy these into your interview, agent, review, or working document. They are specific to this repetition.
template
Layered context pack
Keep these layers separate so they can change at different speeds.
1. OBJECTIVE — task and audience 2. OUTPUT CONTRACT — schema and required fields 3. DURABLE RULES — definitions, constraints, forbidden moves 4. QUALITY BOUNDARY — paired good/bad examples with reasons 5. TASK CONTEXT — decision, source, current constraints 6. RETRIEVED EVIDENCE — only relevant excerpts 7. VERIFICATION — scorer, thresholds, human review
checklist
Context pruning test
For every block, demand a reason it deserves attention.
Which error does this block prevent? Is it durable or task-specific? Could a schema replace prose? Could one paired example replace a paragraph? Is the source authoritative and current? What happens if this block is removed? Can the output cite which context it used?
Calibrate judgment
Compare the evidence, not the polish.
Less context, better fidelity
A transcript-to-evidence task is tested before and after a structured pack.
A schema, no-inference rule, and one paired example reduced unsupported claims from seven to one and review time from 28 to 11 minutes. Removing company history had no negative effect, so it stays out.
Why it works: The task and input stay constant, defects are counted, and context is pruned rather than celebrated by volume.
Context as a document dump
Added the company wiki, strategy deck, all research, brand guide, and a long role prompt. The answer sounded more informed.
Why it fails: There is no task contract, salience strategy, controlled comparison, defect measure, or evidence that the extra material caused improvement.
Review → revise → repeat
The artifact is the beginning of the rep.
Check only standards your current artifact actually meets. Then record one consequential revision before exporting it.