agent-coach

agent-coach is a skill for Claude Code, Codex from contextstream/mcp-server. It costs 97 tokens per session (2,002 once invoked), scanned A, original, MIT.

A system for teaching an assistant to handle recurring work, such as weekly notes, articles, recaps, or image series. It stores workflows, drafts, feedback, and lessons in a ContextStream account.

In plain words
What is it for?
Use it to create recurring workflows, start and close jobs, save drafts, compare versions, record feedback and decisions, and review workflow history or metrics.
Why use it?
It preserves useful feedback between jobs, so repeated work can improve without treating every draft as a fresh task.

Skill for Claude CodeCodex

Written for no agent in particular: nothing here depends on one.

Good fit Use it to create recurring workflows, start and close jobs, save drafts, compare versions, record feedback and decisions, and review workflow history or metrics.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/contextstream/mcp-server/agent-coach
Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Any agent
npx skills add contextstream/mcp-server --skill agent-coach
Clone the repo
git clone --depth 1 https://github.com/contextstream/mcp-server

Made for: Claude Code, Codex.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for agent-coach

README.md
[![agentmods](https://agentmods.dev/badge/skills/contextstream/mcp-server/agent-coach/github.svg)](https://agentmods.dev/skills/contextstream/mcp-server/agent-coach)
Your own site
<a href="https://agentmods.dev/skills/contextstream/mcp-server/agent-coach"><img src="https://agentmods.dev/badge/skills/contextstream/mcp-server/agent-coach/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for agent-coach

Your own site · 80×15
<a href="https://agentmods.dev/skills/contextstream/mcp-server/agent-coach"><img src="https://agentmods.dev/badge/skills/contextstream/mcp-server/agent-coach.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 97 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 2,002 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00097 $0.02002
Opus 5.5 $0.00039 $0.00801
Sonnet 5.5 $0.00019 $0.00400
Haiku 4.5 $0.00010 $0.00200

Measured 2d ago against content hash 1e8b3add7af8, method: parsed. Prices are Anthropic first-party input rates as of 2026-10-07, from the pricing page.

Security

Grade A, and why

agent-coach scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

plugins/agent-coach-kit/skills/agent-coach/SKILL.md · 88 lines

How it starts

The opening of the file, as written. The whole thing — 88 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Agent Coach Kit

The person is teaching you a piece of work they need again and again. Each time, do the work well, make giving feedback easy, and carry what they teach you into the next job.

Words

  • Workflow: the recurring work (weekly-note). It has a goal: what good looks like every time.
  • Job: one occasion of the work (weekly-note#4). It has an objective: what this one must achieve.
  • Version: each draft you save (weekly-note#4/v2).
  • Feedback: the person's own words about one version, kept exactly as said (weekly-note#4/f1). It is either reusable (later jobs too) or this job only.
  • Guidance: your reading of reusable feedback (weekly-note/g3). The person can edit it, limit where it applies, retire it or remove it.

Everything lives in the person's ContextStream account. Three tools reach it:

  • workflow reads: list, get, context, compare, history, metrics, export.
  • workflow_save saves: create, start_job, save_output, add_feedback, record_verdict, close_job and the changes below.
  • workflow_delete deletes one whole workflow, only after the person confirms.

Each job

  1. Find the workflow with workflow list.
    • For new work, help the person turn a rough idea into a short goal: who it's for, the result they want, an example or two, and any limits.
    • Show the goal you propose in one line and ask only about what's missing. Then save it with workflow_save create (name, goal).
    • A rough goal is fine; it gets sharper as you go.
  2. Propose the objective.
    • workflow get shows recent objectives, the guidance, what still needs attention in an open job, and any coaching preference the person set. Feedback it shows from earlier jobs is history: never apply it.
    • Show the objective you propose in one line, and ask only about what changed or is missing. If the person asked for quick coaching, state it in that line and go on.
    • Then workflow_save start_job (workflow, objective). To pick up an open job, read workflow context for it.
  3. Use what you were given. workflow context holds the goal, the active guidance, this job's own feedback, and approved examples. Each guidance item shows what the person actually said beside your earlier reading, plus any limit on where it applies.
    • Apply each guidance item that fits this job. If its limit says it doesn't apply here, say so.
    • Approved examples show what worked. Learn from them, but don't copy them: a different result can be just as good.
    • Feedback from other jobs is left out on purpose. Don't bring back one-off choices you remember from them. In a comparison, anything marked "this job only; not for later jobs" stays with its job.
  4. Save the draft before you show it with workflow_save save_output (job, and text, url or the attached file).
    • A revision: add revises with the version it revises.
    • Pass the context_etag from the context you wrote from.
    • carry_note says briefly what you kept from approvals and which feedback you applied.
    • Name the same items in carried, one each: {"ref": "weekly-note/g1", "how": "applied"}. The how is kept, adapted, applied or not_applied. Items marked applied or kept stop showing as needing attention.
  5. Show it with its version reference. In a line or two, say what changed since the last version and why, and what still needs attention.
  6. Record each reaction. When the person reacts to a version, record their words exactly with workflow_save add_feedback (output, kind correction, partial_approval or direction, scope workflow or job, words, and your interpretation kept apart).
    • Recording the reaction is part of revising; tell them what you recorded rather than asking first whether to remember it.
    • Choose the scope from what they said, and say which in one line: "Saved for later jobs too; say 'just this job' if not." Ask only when a wrong guess would matter.
    • To change the scope later: workflow_save rescope_feedback.
    • When they approve one part, add excerpt. For audio or video, add time_start_ms and time_end_ms.
    • When they repeat something already in the guidance, add repeat_of with that guidance. That shows the coaching didn't stick.
  7. Revise or finish.
    • When they're happy: workflow_save close_job (job, used).
    • To compare with the last job: workflow compare, then ask which is better.
    • As soon as they say, record their verdict with their words: workflow_save record_verdict (job, verdict better, same or worse, note with their words). It works before or after the job is closed.
    • The verdict is the person's choice.

Read the full file on GitHub · 88 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 88 lines · 97 tokens per session scan A 1e8b3add7af8

Subscribe to this mod's changes

agent-coach is a skill published in the GitHub repository contextstream/mcp-server (44 stars, last pushed yesterday), licensed MIT. It adds 97 tokens to every session and 2,002 once invoked, about $0.0004 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-10-14.