Pick Tools That Are Just Right + Stop Context Bloat

What these notes are

The previous lesson covered how to write. This one moves up two layers: what you pick to do the work (L3, tools) and how to stop the context from growing without end (L4, context).

A quick spoiler: the real bulk sits in the context layer.

L3 · Model and tool layer: not the strongest, but just right

Model tiering: read the task first, then pick the tier

Script first: if a machine can compute it deterministically, do not let the model read and then compute

Lookups, filtering, calculations and caching all belong here. Asking a model to count a column or total a table is slow, expensive and error-prone; a script does it in one line.

The four tool principles

Minimal install · enable on demand · lazy load · exit unrelated tools when the task ends

Tools, MCPs and Skills put names, descriptions and parameter schemas into the context. The more tools and the more complex their definitions, the higher the potential cost — even if you never call them.

RTK: filter before CLI output enters the context

RTK (Rust Token Killer) filters and compresses command-line output before it enters the LLM context.

Check whether it is installed first:

rtk --version
rtk gain

Once installed, rtk gain shows the savings statistics. The project publishes benchmarks showing large reductions for common development commands, but real results vary by project, command and workflow.

Raw data → machine filter / compress → key results → LLM judgment

Demo 3 · Tool output compression: teach the Agent to eat less log

A. Manual filtering (the simplest option — learn this one first)

npm test 2>&1 | tail -n 100
grep -iE "error|failed|exception" app.log | tail -n 100

B. RTK (optional, for AI coding scenarios)

See the previous section. Once installed, use rtk gain to compare the savings.

What to watch: compare the length before and after filtering, then check whether the model's judgment got worse.

L4 · Context layer: the real bulk usually sits here

8.1 One task, one session

One goal → one session → done → start a new session for the next task

Do not keep mixing unrelated tasks into the same context. Code, deployment, weekly reports and marketing plans in one session means every round pays for all the history before it — and the AI gets pulled off course by stale topics.

8.2 Compress long sessions into a handoff summary

Compress the current session into the minimum context needed to keep working.

Keep only:
1. Current goal
2. Confirmed facts
3. Completed items
4. Current approach / key decisions
5. Open issues
6. Next step
7. File names, variable names, interface names and commands that must be preserved

Delete:
- Pleasantries
- Rejected options
- Repeated explanations
- Intermediate trial and error
- Anything unrelated to what comes next

Requirement: a new session must be able to continue from this summary alone.

8.3 Fetch files on demand, do not stuff in the whole book

Search first → find the relevant section or page → read only that → answer

The core rule: take what you need, not everything you have.

8.4 Let scripts handle deterministic filtering

See Demo 3. Never let raw logs appear in the model context.

Demo 2 · Context slimming: from a long conversation to a minimal handoff summary

Demo 1 was a diet for a prompt; this one is a diet for a conversation. Use this extended prompt to make the old session produce its own handoff summary:

Compress the current project into the minimum context needed to keep working in a new session.

Must keep:
- Goal
- Confirmed requirements
- Current architecture / approach
- Completed work
- Open issues
- Key names of files, interfaces, variables
- Next step

Must delete:
- Pleasantries
- Repeated explanations
- Rejected options
- Full debugging process for already-solved errors
- History unrelated to the next step

Finally add:
[Suggested next prompt]
so I can copy it into the new session and continue directly.

Steps

Three things to watch: whether the input got noticeably shorter; whether the AI still knows the goal, the key decisions and the next step; whether the noise from old topics is gone.

Advanced idea 1 · Keep raw data out of the context

One family of tools takes the approach that raw data should never enter the context at all — only a distilled result comes back. Typical benefits:

The scenarios that fit all share one trait: large input, small conclusion — Playwright page captures, CSV statistics, log triage, long-document extraction.

Advanced idea 2 · Stop the Agent from digging through the repo every time

This is the most typical waste in AI coding scenarios.

Common capabilities: search · callers · callees · impact · context · affected.

When it is worth the most

• Fixing bugs in an old project: check callers, callees and impact before touching anything, so you stop acting on a local view alone;
• Asking questions about a large repo: questions like how does the order flow work or where does the cache expire are exactly what structural exploration answers;
• CI test selection: use affected as the first filter, which suits repos with many tests and slow full runs.

Lesson recap

The next lesson turns these moves into daily habits and hands you a 30-second checklist.