You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
# Editing Copilot content instruction and agent files
6
6
7
-
This applies when you add, edit, or remove a Copilot instruction or shared agent file that guides how **content** (articles, data files) is written. It does **not** apply to code instructions or agents owned by the engineering team (for example `code.instructions.md`).
7
+
This applies when you add, edit, review, or remove a Copilot instruction or shared agent file that guides how **content** (articles, data files) is written. It does **not** apply to code instructions or agents owned by the engineering team (for example `code.instructions.md`).
8
8
9
-
When it applies, you **must** first read the instruction-architecture doc in full and follow it:
9
+
Content-writing guidance goes in one of two places. Decide which before writing anything.
Read the current version every time (from a local `github/technical-content` checkout if you have one, otherwise fetch the URL); do not rely on your memory of it, because it changes. If you cannot access it, say so and stop rather than guessing.
13
+
These load automatically into every Copilot interaction whose files match the `applyTo` pattern. Use them only for rules that apply to every content interaction, regardless of persona, task, or phase.
14
+
15
+
* Keep the always-on set that loads for content under roughly **150 discrete rules**, with a soft backstop of **~6,500 tokens**. Rule count matters more, because adherence slips as the number of instructions grows. Both are soft warnings: slightly over is fine. Run `npm run measure-instruction-budget` to see where a change lands, and report the before and after numbers.
16
+
* Front-load the most important rules. Rules earlier in a file are followed more reliably, and the order in which separate files are combined is not guaranteed.
17
+
* A rule earns a spot only if it scores well on these factors:
18
+
***Model default gap**: The model consistently gets it wrong without the rule.
19
+
***Breadth**: It applies to every content interaction, not just some personas, tasks, or phases.
20
+
***Cost of missing it**: Getting it wrong causes expensive rework (wrong audience, wrong scope), not polish that is cheap to catch in review.
21
+
***Conciseness**: It fits in one or two short, actionable rules.
22
+
* Write rules as short imperatives, not multi-paragraph explanations.
23
+
24
+
## Shared agent files (`.github/agents/`)
25
+
26
+
Writers invoke these deliberately. Each agent has its own budget (up to 30,000 characters), so guidance that is narrow, detailed, or needs room to explain goes here. There are three kinds:
27
+
28
+
***Persona agents** target a specific audience: `builder-writer` (developers building software) and `driver-writer` (admins and decision-makers enabling developers at scale).
29
+
***Task agents** help with a specific type of work, such as a content design plan or release notes.
30
+
***Review agents** give a focused editing pass on existing content, such as readability.
31
+
32
+
Use agents, not skills, for docs-writing guidance. Skills can be invoked automatically and unpredictably, which undermines consistent output across the team.
33
+
34
+
## Always
35
+
36
+
* Do not duplicate guidance that already exists in another instruction file or shared agent. Check before adding.
37
+
* Keep content-writing guidance in these shared files, never in personal instructions, so it stays consistent across the team.
38
+
39
+
The reasoning behind these rules is written up for humans at https://github.com/github/technical-content/blob/main/contributing-to-docs/docs-work/copilot-instruction-architecture.md. You do not need to read it to follow them.
# Detect PRs that only changed package-lock.json (no engineering source files).
54
52
# These are usually cross-platform `npm install` churn from contributors
55
-
# editing content. We comment with reset instructions instead of pulling in
56
-
#docs-engineering for review.
53
+
# editing content. We comment with reset instructions instead of adding
54
+
#them to the engineering review board.
57
55
#
58
56
# Dependabot is exempt. Its security updates for transitive dependencies
59
57
# change only the lockfile, because the dependency is not in package.json.
60
58
# Those PRs are intentional, so the reset instructions are wrong and
61
-
#suppressing the review request leaves them with no reviewer at all.
59
+
#keeping them off the board leaves them without engineering triage.
62
60
- name: Detect lockfile-only churn
63
61
id: detect
64
62
env:
@@ -97,15 +95,19 @@ jobs:
97
95
git push
98
96
```
99
97
100
-
If the lockfile change is intentional, remove the `lockfile-churn-only` label and request docs-engineering review.
98
+
If the lockfile change is intentional, remove the `lockfile-churn-only` label and push or reopen the PR so this workflow sends it to the engineering review board.
Copy file name to clipboardExpand all lines: content/copilot/reference/copilot-billing/models-and-pricing.md
+1-1Lines changed: 1 addition & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -33,7 +33,7 @@ All prices are **per 1 million tokens**.
33
33
34
34
{% data reusables.copilot.extended-context-pricing %}
35
35
36
-
{% data variables.copilot.copilot_gpt_56_sol %}, {% data variables.copilot.copilot_gpt_56_terra %}, {% data variables.copilot.copilot_gpt_56_luna %}, {% data variables.copilot.copilot_gpt_6_astra %}, {% data variables.copilot.copilot_gpt_6_luna %}, and {% data variables.copilot.copilot_gpt_6_sol %} include a cache write cost in addition to cached input. Earlier OpenAI models have no cache write cost.
36
+
{% data variables.copilot.copilot_gpt_56_sol %}, {% data variables.copilot.copilot_gpt_56_terra %}, {% data variables.copilot.copilot_gpt_56_luna %}, {% data variables.copilot.copilot_gpt_6_astra %}, {% data variables.copilot.copilot_gpt_6_luna %}, {% data variables.copilot.copilot_gpt_6_sol %}, and {% data variables.copilot.copilot_gpt_61_sol %} include a cache write cost in addition to cached input. Earlier OpenAI models have no cache write cost.
0 commit comments