improve-agent
Use when Improve should run in a fresh native agent instead of this conversation: start one agent that runs /improve in its own context, keep useful parent…
Skill by whichguy
Chalk Labs
Need help deploying improve-agent?
Chalk Labs sets up agents like this inside real teams: connected to your tools, tested on your data, with a human in the loop where it matters.
Install
Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.
npx agentmarket install skill/whichguy-improve-agent
Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: agent-building
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
Permissions
What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.
- Tools
- No tools found
- Secrets
- None asked for
- Install scripts
- None found
Network hosts it talks to 1
codeload.github.com· found in install command
Files it touches
~/.claude/skills/improve-agent
What's inside
- Skill name
improve-agent- Description
- Use when Improve should run in a fresh native agent instead of this conversation: start one agent that runs /improve in its own context, keep useful parent work going, then verify and relay its result. For Improve in this conversation, use improve.
- License
- MIT
- Metadata
- short-description: Run Improve in a fresh native agent
- Instructions
- 139 lines in SKILL.md · view on GitHub
Files (4)
- LICENSE
- README.md
- SKILL.md
- openai.yaml · agents/
SKILL.md (first lines)
# improve-agent
This card delegates one Improve invocation to one fresh native agent. The
selected `improve` card owns the review-and-improvement algorithm and runs
inline wherever it is invoked; this card owns only the dispatch, the assignment,
the worker contract and the parent's verification. The host-selected Ask Agent
card owns native launch, workspace preparation and delivery mechanics.
Do not run the Improve loop in this conversation, and never start a second
Improve writer for the same candidate. When the user wants Improve without an
agent, use `improve` instead.
## Resolve the selected cards and host
Resolve the absolute paths of the **selected, loaded** `improve/SKILL.md` and
Ask Agent `SKILL.md` supplied by the host, expanding any selected skill-root
alias. Do not guess a sibling checkout, an older installed wrapper, `PATH`, or a
same-named skill. Read the Ask Agent card's route selection, current-learnings,
approval and result sections before launch.
Check the live host schema, following Ask Agent's host-capabilities reference:
the agent needs fresh context without the inherited conversation, access to the
workspace and the selected skills, the tools the task needs, and a native result
collection route. If a card or a required capability is missing, report the gap
and stop. Do not substitute inherited context, a shell-launched model, an
external runner, or an inline run of Improve; running inline instead is the
user's choice.
## Choose the workspace route
- **Standalone request.** Select `workspace_route: helper-managed` explicitly;
it is not Ask Agent's default. Its
helper prepares an isolated worktree from the current candidate snapshot.
Choose the delivery mode before launch: `commits` when Improve's ordinary
commit policy applies, `patch` for an explicit no-commit request, and
`report-only` for a preview or dry-run request.
- **Consumer-bound child.** When a consumer such as a ShipLoop
`delegation: ask-agent` Improve packet supplies a complete consumer-owned
contract, use Ask Agent's consumer-owned workspace route with
`workspace_route: consumer-owned` and `delivery_mode: in-place`. It requires
the selected Ask Agent card's `ask-agent/consumer-owned-workspace/v1`
capability. The parent writes the consumer's `host-owner.md` record as that
route describes. If the capability or any contract field is missing, fail
closed: keep the action pending with its locators and never fall back to the
helper-managed route or a second worktree.
## Compose the assignment
Write the native assignment in this order, keeping essential facts and actual
decisions inline with locators for supporting detail. The worker must not need
this card, a whole parent packet or a consumer reference to begin.
1. **Current context and desired improvements**, with **Current learnings**
nested inside it, following Ask Agent's current-learnings rule: what is known,
what already changed, and what should be improved or investigated next.
Keep facts, decisions and hypotheses visibly distinct; suggested fixes are
candidates to assess, not findings or authorization.
2. `Run /improve using <selected absolute improve SKILL.md>`, once.
3. The binding:
- the workspace's absolute path and Git root, and the route markers;
- the candidate scope or inventory, its exclusions, and the base or HEAD;
- authority: commit, no-commit and push rules, the current approvals,
declines and pending decisions with their sources and conditions, and any
explicit independent-review request (without one, the worker's Improve
run keeps its inline self-review default);
- the evidence root and, for a consumer-bound child, the exact child packet
receipt, reviews directory and `host-owner.md` locators and the binding
marker the consumer printed;
- the return route in [Worker contract](#worker-contract), and the
parent-only continuation it must leave unexecuted.
A consumer-bound assignment also carries these literal markers after the
opening:
```text
workspace_route: consumer-owned
delivery_mode: in-place
execution_role: improve-executor
delegation_owner: parent
```
## Worker contract
Give the worker these rules in the assignment:
- Verify the process `pwd` and `git rev-parse --show-toplevel` against the
assigned workspace before task work and after any directory change, and use
that workspace as the operation directory for every shell call. A mismatch is
a blocker to return, not something to work around.
- Run the selected Improve card once, inline in your own context, under the
frozen authority. Its standalone binding, including any ShipLoop whole-skill
subcall section the assignment names, applies unchanged.
- Be the only candidate writer. Do not dispatch Improve again, create another
worktree, or start reviewer, test-runner or executor agents unless the
assignment carries an explicit independent-review request. Collect any
delegate before returning.
- Do not execute a parent callback, workspace return, merge, push beyond the
stated authority, caller delivery or cleanup. Read `host-owner.md` for
orientation only; the parent alone appends to it.
- Save each runtime packet at the given receipt path, finish all writes, and
return inline:
- Improve's completion summary: key implemented changes, what was learned,
validation and remaining work;
- the observed workspace and Git root, initial and final HEAD, changed paths,
and scoped commit SHAs or the explicit no-commit or no-change reason;
- the checks run and the review-record locators, and the terminal packet and
completion-evidence locators;
- the runtime's terminal status (`complete`, or the incomplete `stopped` or
`blocked` state with its blocker), and a statement that no parent-owned
step was executed.
## While the agent runs
Continue useful parent work that neither writes the candidate nor runs
input-mutating checks in it. Forward a later user decision to the existing
worker through native messaging; for a consumer-bound child, also append its
source, scope and receipt status to `host-owner.md`. A lost, running or unknown
worker blocks replacement until it is collected or the host shows it stopped or
never launched, as Ask Agent's recovery rules describe.From whichguy/skill-craft, MIT.
More from whichguy/skill-craft
- 0.592score
rubric-eval
whichguy
Evaluate a prompt change against a rubric over a scenario catalog: run arms on a subject model (Grok by default), grade blind with one evidence-first judge…
SkillAFreshness: A★ 1Claude Code +9 - 0.592score
improve
whichguy
Use when a repository candidate needs a deliberate review-and-improvement loop: use recent Git history, make warranted changes, run meaningful checks, and…
SkillAFreshness: A★ 1Claude Code +9 - 0.592score
plan-dispatcher
whichguy
Use when executing or resuming an agreed dependency graph with one main dispatcher and parallel native workers: claim ready steps, preserve launch state…
SkillAFreshness: A★ 1Claude Code +9 - 0.592score
adversarial-review
whichguy
Have an independent model try to break a proposed change and return its findings as testable hypotheses, each with the smallest experiment that would refute it.
SkillAFreshness: A★ 1Claude Code +9 - 0.592score
review-coverage
whichguy
Add a post-ship improve-to-exhaustion directive to a plan, or run that directive after implementation.
SkillAFreshness: A★ 1Claude Code +9 - 0.589score
in-parallel
whichguy
Use for "in parallel" or when one goal should be divided among independent native agents and returned as one verified aggregate.
SkillAFreshness: A★ 1Claude Code +9 - 0.589score
ask-agent
whichguy
A delegation skill, not an agent type.
SkillAFreshness: A★ 1Claude Code +9 - 0.588score
shiploop-e2e-audit
whichguy
Run the ShipLoop test harness and audit its retained graph, review, test, product and incremental-change evidence.
SkillAFreshness: A★ 1Claude Code +9 - 0.588score
backchain
whichguy
Use when an implementation task needs a dependency-aware plan before coding: backward planner / backchain / precondition-first planning, dependency DAGs…
SkillAFreshness: A★ 1Claude Code +9 - 0.587score
architect
whichguy
Design system architecture and make technology decisions.
SkillAFreshness: A★ 1Claude Code +9 - 0.587score
plan-test
whichguy
Generate comprehensive tests for code.
SkillAFreshness: A★ 1Claude Code +9 - 0.587score
improve-system-prompt
whichguy
Benchmark and compare system prompt variants (V2/V2a/V2b/V2c) for Sheets Chat by running test scenarios through the real GAS-side ClaudeConversation pipeline.
SkillAFreshness: A★ 1Claude Code +9
Signals
- Stars
- 1
- Forks
- 0
- Open issues
- 0
- Last commit (this folder)
- Sep 28, 2026 (1 month ago)
- Last release
- Aug 15, 2026 (2 months ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- MIT
- Repository
- github.com/whichguy/skill-craft
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from whichguy
- 0.592score
plan-dispatcher
whichguy
Use when executing or resuming an agreed dependency graph with one main dispatcher and parallel native workers: claim ready steps, preserve launch state…
SkillAFreshness: A★ 1Claude Code +9 - 0.592score
rubric-eval
whichguy
Evaluate a prompt change against a rubric over a scenario catalog: run arms on a subject model (Grok by default), grade blind with one evidence-first judge…
SkillAFreshness: A★ 1Claude Code +9 - 0.592score
adversarial-review
whichguy
Have an independent model try to break a proposed change and return its findings as testable hypotheses, each with the smallest experiment that would refute it.
SkillAFreshness: A★ 1Claude Code +9 - 0.592score
improve
whichguy
Use when a repository candidate needs a deliberate review-and-improvement loop: use recent Git history, make warranted changes, run meaningful checks, and…
SkillAFreshness: A★ 1Claude Code +9 - 0.592score
review-coverage
whichguy
Add a post-ship improve-to-exhaustion directive to a plan, or run that directive after implementation.
SkillAFreshness: A★ 1Claude Code +9 - 0.589score
ask-agent
whichguy
A delegation skill, not an agent type.
SkillAFreshness: A★ 1Claude Code +9
Done for you
Rather not set it up yourself?
Chalk Labs scopes, builds and runs agents for your team: the task audit, the eval harness, the integrations and the weekly ops report. Fixed scope, priced before we start.
See how we build AI agents →Read the case studies
Chalk Labs is a two-founder AI and growth studio. We build AI agents, CRMs and MVPs for Web3 and AI companies, and have shipped 50+ product and token launches.
- launches shipped
- 50+
- launches shipped
- media placements
- 500+
- media placements
- founders on every account
- 2
- founders on every account
Press we've placed: Forbes · CoinDesk · Decrypt · Cointelegraph · The Block
Rahil Jain, Co-founder, Chalk Labs. Ex-Goldman Sachs engineer. You talk to the person who builds it.
Get improve-agent running in your business
Tell us what you want to automate. A founder replies within one working day with a plan, not a pitch deck.
Cite this page
improve-agent skill. Agent Mart. https://agentmart.dev/skills/whichguy-improve-agent (last updated 2026-10-24)