EC1B1 Python Coding SupportFollow-up resources for LSE BSc Economics
← Back to resources

Session 4 Toolkit: Planning, Skills, and Agent Orchestration

Narrative guide to applying Session 4 methods: planning mode, reusable skills, sub-agents, and anchor files.

Updated 2026-02-13

Why this page exists

Session 4 marked a clear shift in how we use AI coding tools. The point was not to collect more commands or to find the fastest way to generate code. The point was to design a reliable workflow: one where planning happens before execution, where repeated tasks are standardized, and where consistency is checked systematically.

If you only remember one idea from this session, remember this: when your process is clear, your prompts become shorter, your output becomes cleaner, and your review time becomes manageable.

The Session 4 pattern in plain language

In the live demo, the same principle appeared repeatedly. First, define the project and constraints. Second, ask for options and a plan. Third, decide deliberately. Fourth, execute in controlled chunks. Finally, review for internal consistency.

That pattern is simple, but it prevents the most common failure mode in agentic coding: getting a lot of output quickly and spending far too long trying to untangle it afterwards.

A practical weekly workflow for EC1B1

Before you ask any agent to write code

Write a short project brief in your own words. Include your research question, data choices you are considering, and the output you want to produce this week. This can be one page; it does not need to be perfect.

Then run planning mode and compare alternatives. Make choices explicitly, especially where economics context matters (frequency, variable definitions, sample periods, and assumptions).

While agents are executing tasks

Keep tasks narrow. Ask for one step at a time and review quickly after each step. If you spot a mismatch, correct it immediately rather than letting it propagate across files.

At this stage, reusable skills save a lot of effort because they standardize how you ask recurring tasks (literature review, data pulls, transformations, diagnostics).

Before you call the work finished

Run a consistency pass. Check whether your literature notes, data strategy, and empirical implementation still align. This is where sub-agents and anchor files are most valuable.

Key references by purpose

If you need to understand agent modes and workflow control, start with Cursor docs on Agent modes.

If you want to build repeatable prompts, read Cursor Skills and Claude Code Skills.

If your project has multiple moving parts and you need consistency checks, read Cursor Subagents and Claude Code Sub-agents.

If you need persistent project constraints, read Cursor Rules and CLAUDE.md guidance.

Companion pages on this site

For tool-specific notes and curated links, use these pages:

If your project starts drifting

When your folder starts to feel inconsistent, do a short reset instead of continuing blindly:

  1. Re-read your project brief and anchor file.
  2. Ask for a fresh plan based on current files.
  3. Compare the plan against what has actually been implemented.
  4. Create a fix list and resolve it in small steps.

This reset is often faster than patching a large amount of unstructured output.

Session links