Project Context Kit

Give your agent a memory that survives between sessions

The single biggest upgrade to working with an agent isn’t a better prompt — it’s persistent context. Instead of re-explaining your project every time, you keep two small Markdown files the agent reads at the start of a session and updates at the end:

• A HUB — the standing state of the project: what it is, where the data and code live, who’s waiting on what, the next few steps, and the open questions.
• A session log — the running history: what each work session did, the decisions you made and why, and a “start here next time” handoff.

The HUB answers “where do things stand?” The log answers “how did we get here?” Together they let a fresh session pick up where the last one left off.

Below is a worked example for a fictional study, plus a prompt you can paste into your own agent to build the same thing for your project.

The sample project — Cash Transfers & School Attendance (Kenya) — is fictional. Names, paths, and results are invented to show the format, not a real study.

Build your own (copy-paste prompt)

This is the actionable part. Edit the bracketed bits and paste it into Claude Code, Claude, ChatGPT — anything that can see your project folder. It creates the two files for your project, then keeps them current.

Download the prompt →

Step 1 — create the two files (run once):

You're helping me set up persistent context for a research project so that you (and future
sessions) can pick up where we left off without me re-explaining everything.

My project: [one or two sentences — what it is and the deliverable].
Project folder: [path, e.g. ~/Dropbox/GitHub/my-project/ — or describe where things live].

Please create two Markdown files in [folder]:

1. HUB.md — the standing state of the project. Include:
   - Status / stage / last-updated / owner
   - A "Where Everything Lives" table (data, code, draft, IRB, pre-registration — whatever applies)
   - Team & human dependencies (who's waiting on what, with dates)
   - Current state (done / in progress / blocked)
   - Next 3 steps, each concrete enough to act on
   - Key decisions & open questions
   - A status log table at the bottom

2. session-log.md — a running, newest-on-top history. For each work session, record:
   - Date and a one-line goal
   - What happened
   - Decisions made, each with the *why*
   - Files touched
   - A "Start here next time" handoff line

Before you write anything, look through my project folder and ask me up to 5 questions about
anything you can't infer (team, deadlines, where the data lives, open decisions). Then draft both
files and show them to me for edits.

Step 2 — keep them current (end of each session):

Before we stop: update HUB.md and session-log.md to reflect today's work.
- Add a new session-log entry on top (goal, what happened, decisions + why, files touched,
  and a "Start here next time" line).
- Update the HUB's current-state, next-3-steps, open-questions, and status-log so they match reality.
Keep it concise — these are working notes, not a report.

Step 3 — rehydrate at the start of a later session:

Read HUB.md and session-log.md before we begin, then tell me in 3–4 bullets where the project
stands and what the "start here next time" note said. Then we'll pick the next step.

Sample 1 — the HUB

The standing state of the project. Keep it to about a page; if it’s painful to maintain, it won’t get maintained.

View / download sample-HUB.md →

Show the full sample HUB

# Cash Transfers & School Attendance (Kenya)

**Status**: Active
**Stage**: Analysis
**Last Updated**: 2026-06-12
**Owner**: A. Researcher

## Quick Summary

A clustered RCT in 120 primary schools testing whether unconditional cash transfers to
households raise daughters' school attendance. Deliverable: a working paper + replication
package by Fall 2026. We are mid-analysis on the midline data.

## Where Everything Lives

| Resource | Service | Path / Link |
|----------|---------|-------------|
| **Raw data** | Box | `Box/CashTransfers_KE/01_raw/` |
| **Working data** | Box | `Box/CashTransfers_KE/02_working/` |
| **Output (tables/figures)** | Box | `Box/CashTransfers_KE/03_output/` |
| **Code** | GitHub | `github.com/aresearcher/cash-transfers-ke` |
| **Paper draft** | Overleaf | "Cash Transfers KE — working paper" |
| **IRB** | UVM IRB | Protocol #STUDY00004821 (approved 2025-03) |
| **Pre-registration** | AEA RCT Registry | AEARCTR-00xxxxx |

## Team & Human Dependencies

| Person | Role | Waiting on them? | They waiting on you? | Deadline |
|--------|------|-------------------|----------------------|----------|
| A. Researcher | PI | No | — | — |
| J. Co-author | Co-PI (analysis) | Attendance recode decision | No | Jun 20 |
| Field RA (Nairobi) | Data | Cleaned endline roster | No | Jul 1 |

## Current State

- **What's done**: Baseline + midline cleaned; balance table passing; main attendance regression running.
- **What's in progress**: Heterogeneity by household poverty; how to handle students who transferred schools mid-year.
- **What's blocked**: Endline data not yet delivered from the field (expected ~Jul 1).

## Next 3 Steps

1. [ ] Resolve the transfer-student recode with J. Co-author (affects the attendance denominator). Due Jun 20.
2. [ ] Rebuild the main results table once the recode is settled.
3. [ ] Draft the heterogeneity section from the poverty-interaction results.

## Key Decisions & Open Questions

- **Open**: Do students who transfer schools mid-year count as "attending"? Two defensible codings; results move ~2pp. Need J. Co-author's call.
- **Decided (2026-05-30)**: Cluster SEs at the school level (the unit of randomization), not the household.

## Notes

- Randomization is at the **school** level — always cluster SEs there.
- `01_raw/` is read-only. All cleaning writes to `02_working/`.
- Paired with the session log — this HUB holds current state; the log holds how we got here.

## Status Log

| Date | Update | Session |
|------|--------|---------|
| 2026-06-12 | Balance table passing; started heterogeneity by poverty | S4 |
| 2026-06-05 | Main attendance spec finalized; SE clustering decided | S3 |
| 2026-05-28 | Midline data cleaned and merged to baseline | S2 |
| 2026-05-20 | Project HUB created; file locations mapped | S1 |

Sample 2 — the session log

The running history. Newest entry on top. Each entry captures the goal, what happened, decisions and the reasoning behind them, files touched, and where to start next time.

View / download sample-session-log.md →

Show the full sample session log

# Session Log — Cash Transfers & School Attendance (Kenya)

## 2026-06-12 · Session 4 — Heterogeneity by poverty

**Goal:** Start the heterogeneity analysis now that the main spec is settled.

**What happened:**
- Confirmed the balance table still passes after the midline merge (no differential attrition flags).
- Ran the main attendance regression interacted with a baseline poverty index (PCA of asset ownership).
- First pass: effect looks ~twice as large in the poorest tercile, but CIs overlap — needs endline data.

**Decisions:**
- Use the **baseline** poverty index for heterogeneity, not midline — midline poverty is post-treatment
  and would be a bad control. *(Why: avoids conditioning on an outcome.)*

**Files touched:** `code/04_heterogeneity.do` (new); `03_output/f3_het_poverty.png` (draft)

**Start here next time:** Don't finalize heterogeneity numbers until the transfer-student recode is
resolved (see open question in HUB) — it shifts the denominator.

## 2026-06-05 · Session 3 — Main spec + SE clustering

**Goal:** Lock down the main attendance regression so results are stable.

**What happened:**
- Specified the primary model: attendance on treatment, school strata FE, baseline controls.
- Reviewed whether to cluster at school or household.

**Decisions:**
- **Cluster SEs at the school level** — the unit of randomization. *(Why: clustering below the
  randomization unit understates SEs.)*

**Files touched:** `code/03_main_results.do`; `03_output/t2_attendance.tex`

**Start here next time:** Begin heterogeneity by poverty. Build the baseline poverty index first.

## 2026-05-28 · Session 2 — Clean and merge midline

**Goal:** Get midline into a usable analysis file.

**What happened:**
- Cleaned the midline roster; fixed 14 duplicate student IDs (kept most recent).
- Merged midline to baseline on `student_id`; 3% unmatched, all students who left the sample area.

**Decisions:**
- All cleaning writes to `02_working/`; `01_raw/` stays untouched. *(Why: reproducibility.)*

**Files touched:** `code/02_clean_midline.do`; `02_working/midline_clean.dta`

**Start here next time:** Specify the main attendance regression and settle SE clustering.

## 2026-05-20 · Session 1 — Project setup

**Goal:** Stand up the project so future sessions have context.

**What happened:**
- Created the project HUB: mapped where data, code, and the draft live; listed the team.
- Confirmed IRB and pre-registration IDs.

**Files touched:** HUB.md (new)

**Start here next time:** Clean and merge the midline data.

---

These pair with the June 15 webinar materials. The same idea scales up: as projects accumulate, a folder of HUBs becomes a dashboard of everything you have going on.