Agent Work Artifacts Layout

Agent work artifacts: where do they go?

When to use

Any task where the agent produces files that are not the user's intended deliverable: PR body drafts for gh pr create --body-file, throwaway repro_*.py / check_*.sh scripts, audit reports, analysis dumps, temporary fixtures. Without a convention these end up either polluting the working tree (triggering "do I need to delete this?" round-trips) or getting lost across sessions even when they had reuse value.

Problem

Three recurring failure modes:

1. PR body files left in the repo root (e.g. pr_body.md),

forcing the user to confirm deletion after every PR.

1. One-off scripts committed to random places — sometimes in

the repo root, sometimes nowhere — making future contributors unsure whether they are project tooling or dead weight.

1. Fixed temp paths like /tmp/pr_body.md that collide between

parallel agent sessions.

Solution

Route every artifact through this four-way decision:

1. Single-use (PR body, gh intermediate, shell pipe)?
   → system temp via  mktemp -t .XXXXXX.
     (POSIX: $TMPDIR or /tmp; Windows: %TEMP%)

2. Repo-scoped helper, human-authored, part of project lifecycle?
   → scripts/  (or  tools/  if the repo already uses that)
     with argparse CLI, docstring, usage example.

3. Repo-scoped helper, agent-authored, reviewed and reusable?
   → .agent/tools/  (tracked subdirectory of the agent working area,
     see the four-cell layout below)

4. General-purpose, useful across projects?
   → offer to publish as a GitHub Gist.

5. Agent working notes / drafts / caches NOT meant for sharing?
   → .agent/scratchpad/ | .agent/artifacts/ | .agent/context/
     (gitignored, local-only)

If none apply, delete the file rather than leaving it ambiguous.

.agent/ four-cell layout

.agent/ is not a single bucket — split it by lifetime and tracking intent:

| Subdirectory | Purpose | Tracked? | | --- | --- | --- | | .agent/scratchpad/ | Free-form notes, pseudo-code, drafts. | No | | .agent/artifacts/ | Generated deliverables (exported docs, images, one-off reports). | No | | .agent/tools/ | Reviewed agent-authored helpers for this repo. | Yes | | .agent/context/ | Project index / knowledge-base caches / embeddings. | No |

The tracked subset is surfaced via a whitelist-style .gitignore (see Example). Keeping .agent/tools/ in Git lets genuinely reusable agent-authored helpers survive across sessions without mixing with the project's human-authored scripts/.

scripts/ vs .agent/tools/

They coexist, with clear provenance separation:

• scripts/ (if/when the repo has one) = human-authored core

tooling that's part of the project's normal lifecycle (release, lint, bootstrap).

• .agent/tools/ = helpers written by an agent, reviewed, and

kept because they proved useful. Provenance is tagged at the directory level — no silent promotion to scripts/, only via a dedicated "graduation" PR.

Example

Right way — PR body in system temp

body=$(mktemp -t pr_body.XXXXXX.md)
# write body via the editor tool, targeting "$body"
gh pr create --repo / --base  --head  \
  --title '' --body-file "$body"

Wrong way — PR body in repo root

# Leaves pr_body.md in the working tree after the PR is filed.
# The user then has to approve its deletion in a follow-up turn.
vi pr_body.md
gh pr create --body-file pr_body.md

Right way — commit message via repo-local scratchpad

In an agent terminal, commit flows that chain mktemp + git commit

• rm often trigger human-confirmation prompts on every commit.

.agent/scratchpad/ sidesteps this by removing both mktemp and the trailing rm:

MSG=.agent/scratchpad/commitmsg.$$.txt
printf '%s\n' 'Subject line' '' 'Body paragraph.' > "$MSG"
git commit -F "$MSG"
# .agent/scratchpad/ is gitignored; sweep occasionally.

This applies to commit messages, not PR bodies — PR bodies should still use system temp (see previous example) because they are one-shot and benefit from OS-level cleanup.

See [shell-heredoc-and-multiline-strings](../shell-heredoc-and-multiline-strings/SKILL.md) for when to choose each temp location.

Right way — repo-scoped helper that others might reuse

/scripts/check_doc_consistency.py     # committed, with argparse + docstring

Right way — agent working area with whitelist gitignore

/.agent/
  README.md            # tracked: layout convention for contributors
  scratchpad/          # ignored: notes, drafts
  artifacts/           # ignored: generated deliverables
  context/             # ignored: indexes, caches
  tools/
    README.md          # tracked: intake standard
    .py        # tracked: reviewed agent-authored helpers
    .sh        # tracked

Corresponding .gitignore block (whitelist pattern — blanket ignore first, then un-ignore each tracked path):

/.agent/*
!/.agent/README.md
!/.agent/tools/
/.agent/tools/*
!/.agent/tools/*.py
!/.agent/tools/*.sh
!/.agent/tools/README.md

Verify with git check-ignore -v after editing — it's the only way to be sure reopens line up with blanket ignores.

Intake standard for .agent/tools/

Because helpers here are tracked, they need a quality bar. Every committed script should have:

1. Provenance header naming the agent session.
2. argparse CLI with --help.
3. Dry-run default; --apply / --write to mutate.
4. Module-level docstring with a real usage example.
5. Stdlib-only (or deps the project already has).
6. UTF-8 safe (encoding="utf-8" on file I/O).
7. Introduced in its own dedicated PR, not smuggled in.

Document this in a .agent/tools/README.md at repo setup time.

Pitfalls

• The convention is inert without an actual .gitignore. Writing

this skill, or a .agent/README.md, or just telling the agent to "put drafts in .agent/scratchpad/" does not make those files ignored. Without a committed .gitignore entry they show up as ?? in git status, and a single git add -A (easy to type when rushing a commit) sweeps them into the tree. Verify on day one:

``bash git check-ignore -v .agent/scratchpad/anyfile.txt # expect: .gitignore::.agent/ .agent/scratchpad/anyfile.txt ``

If that command prints nothing, the convention is a lie and the next git add -A will betray it. This skill's own repo shipped without a .gitignore for its first nine PRs — nobody got bitten only because every commit used explicit git add . Don't rely on that.

• .agent/ is convention, not project API. A tracked

.agent/README.md explaining the layout is fine (similar in spirit to .github/'s own README), but nothing in the build, tests, or CI should read from .agent/. If a file needs to run on fresh clones or other contributors' machines, it doesn't belong here — promote it to scripts/ with a dedicated PR.

• Whitelist gitignore is order-sensitive. The blanket

/.agent/* must come before the !-prefixed reopens, otherwise the reopens have no effect. Always verify with git check-ignore -v for both a tracked path and an ignored path before committing.

• Use .agent/scratchpad/ for commit-message drafts, not PR bodies.

Commit messages are produced many times per session and benefit from a fixed local path (no mktemp, no rm, fewer agent-terminal confirmation prompts). PR bodies are one-shot, often contain the final version of release notes, and should stay in system temp so the OS cleans them up — a lingering draft under .agent/scratchpad/ invites confusion about which version made it to GitHub.

• **Do not rely on .agent/scratchpad/, /artifacts/, or

/context/ for anything needed on another machine** — they are gitignored. Only .agent/tools/ survives cloning.

• Windows: /tmp and mktemp are not portable. Prefer

tempfile.gettempdir() from Python when the caller might be on Windows.

• Do not confuse scripts/ with contrib/. In PostgreSQL /

Git, contrib/ means "distributed alongside core but not part of it". Unless the target repo already uses that convention, stick with whichever of scripts/ or tools/ the repo already has.

• **Do not silently "graduate" .agent/tools/ scripts to

scripts/.** Promotion is a conscious decision that warrants a PR of its own (with a new provenance line: "originally authored by agent, audited and graduated"). Silent moves destroy the provenance signal the .agent/tools/ location exists to preserve.

• Industry precedent: CPython, Django, Kubernetes, Rust, LLVM

all ship a top-level scripts/ or tools/. No mainstream project formalises an "AI agent working area" yet; .agent/ is the convention this skill proposes, sitting alongside existing local caches like .pytest_cache/, .mypy_cache/, .idea/.

See also

• [github-pr-via-gh-cli](../github-pr-via-gh-cli/SKILL.md)
• [shell-heredoc-and-multiline-strings](../shell-heredoc-and-multiline-strings/SKILL.md)
• [workspace-path-constraints](../workspace-path-constraints/SKILL.md)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: chen3feng
• Source: chen3feng/agent-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.