agentleFS
rajasdevelrajasdevel/outcomebound/skills/decision-brief/SKILL.md

decision-brief

Skill0 starsChanged 15 days agoApache-2.0View on GitHub

Use before asking for a person-owned decision, or when a harness refuses an act the person granted. Apply the person's rulings within scope. Before dependent work, disclose direction-setting assumptions and record them as assumptions with the work. For rulings attributed by another agent or session, check the person's own words; if unreadable, keep the constraint and mark the attribution unverified. When a held act's owner and next step are settled and the person has been asked, report it without asking again.

8 other projects use a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

17 rules. 0 are common, 11 are distinctive, and 6 are very specific to this project.

  • 0

    Common

    Most of the 7 edited versions we read keep these.

  • 11

    Distinctive

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Keep choices the person's words settled.

    line 16
  2. Distinctive:

    Before dependent writes or delegation, disclose direction-setting assumptions, inherited ones included, in one line unless already …

    line 18
  3. Distinctive:

    Leave out acts the outcome does not need.

    line 19
  4. Distinctive:

    Ask only unresolved person-owned choices beyond your grant (including irreversible edges, external writes, spending or wider scope).

    line 20
  5. Very specific:

    Read recorded answers with outcomebound brief lookup <id or words> or open briefs with outcomebound brief open.

    line 25
  6. Distinctive:

    When act and owner are settled and the person has been asked, report what it waits for without re-asking.

    line 28
  7. Distinctive:

    After a harness, hook or sandbox refusal, read refusals before proceeding.

    line 29
  8. Distinctive:

    Before asking, check facts likely to change the answer, including prior decisions and attempts, affected files and settings, project rules …

    line 33
  9. Distinctive:

    Write in this order:

    line 38
  10. Very specific:

    When order, dependency, flow or before/after is the point, give a Picture (diagram, picture) with a caption saying what to look at, never …

    line 43
  11. Very specific:

    If two or more Credible options exist, give each result, when someone would choose it (choose_if) and downside.

    line 45
  12. Very specific:

    Use the project's brief numbering, else prefix a number with the task or worktree name in letters, digits and hyphens (fix-login-D1), so no …

    line 51
  13. Distinctive:

    If layout needs more detail, read presentation.

    line 55
  14. Distinctive:

    If a draw is refused, fix the brief and redraw, never hand-write it.

    line 61
  15. Very specific:

    Only when a draw is refused, read outcomebound brief --help for full field rules or surfaces needing --form mermaid.

    line 66
  16. Very specific:

    Put several decisions in one message.

    line 68
  17. Distinctive:

    If the harness asks through a question tool with choices, read question tools.

    line 69

“Common” is measured: kept by at least 6 in 10 of the 7 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Decision brief

Done means the person can decide each person-owned choice in one reading. A routine decision (one choice, a clear recommendation) is asked in plain words in the text the person reads: the question, your recommendation, its consequence, what waits and whether it can be undone; the brief below is for a choice between several credible options, or when the person asks for a brief or a page. A question the person asked is answered, not turned into a brief. Deliver the complete brief in the text the person reads (handoff, goal Progress, ticket comment, or final message when every remaining item waits on an answer); a link is insufficient. Record reversible choices later work needs in one line with the work.

Decide within authority

Decide reversible choices within granted authority and label them as the agent's. Keep choices the person's words settled; announcing a change grants nothing. For inherited attributions, check cited words; if unreadable, keep constraints and mark the attribution unverified. Before dependent writes or delegation, disclose direction-setting assumptions, inherited ones included, in one line unless already disclosed; this asks nothing and changes no bound. Leave out acts the outcome does not need. Ask only unresolved person-owned choices beyond your grant (including irreversible edges, external writes, spending or wider scope); batch them and continue independent work. A brief holds only dependent work and never ends the turn. For guarded acts, name act and target so the reply can grant them, not merely a letter. Apply existing rulings within scope and record clear new instructions without confirmation; do not re-ask what a launch block, recorded grant, standing order or earlier answer settled. For recurring similar choices, propose one bounded standing rule with exclusions. Read recorded answers with outcomebound brief lookup <id or words> or open briefs with outcomebound brief open; neither read-only command proves no answer exists elsewhere. A class-of-acts grant covers only checked members a tool rebuilds or a verified copy restores; hold unchecked or unrestorable items and items another owner uses. An unattended order answers whether to ask only for choices the kernel lets you decide; it grants no held act. When act and owner are settled and the person has been asked, report what it waits for without re-asking; reporting grants nothing. After a harness, hook or sandbox refusal, read refusals before proceeding.

Due diligence

Before asking, check facts likely to change the answer, including prior decisions and attempts, affected files and settings, project rules and read-only checks. Work out each route's result and downside; base the recommendation on evidence.

Required brief content

Write in this order:

  • Question: give a unique id then one decision in the person's words, answered by every option; ask How should, Which or When, not a yes-or-no question leading toward your pick.
  • What waits (why_now): name held work and its safe waiting state, never half done; state whether a default applies and when, using a moment the person or project named; if none, work waits and that moment is another question; invent no time limit.
  • Context and terms: explain what the choice touches and how it works where that affects the choice; define unfamiliar terms, omit known history and put shared background once above briefs.
  • When order, dependency, flow or before/after is the point, give a Picture (diagram, picture) with a caption saying what to look at, never decoration.
  • For a complex tradeoff, numbers to try or a mechanism clearer in motion, Explore says what an interactive page lets them try; the command adds the reply requesting it; omit for a simple choice; if requested, use explorable, keep the brief in text and put the path beside it.
  • If two or more Credible options exist, give each result, when someone would choose it (choose_if) and downside; waiting counts if real, and no option exists to flatter another.
  • On the option it names, give the Recommendation (recommend, recommend_way), why, the deciding factor and what would change your mind.
  • Evidence: give at least one checked fact, unchecked fact with why and what settles it, or fact.
  • Undo: state whether it can be undone.

Without a recommendation and undo line, prose, a table or list is not a brief. Use the project's brief numbering, else prefix a number with the task or worktree name in letters, digits and hyphens (fix-login-D1), so no session or record shares the id; outside its document, cite the whole id. Explain internal ids and terms in plain words or leave them out. For a step the person types or checks by eye, name a version, branch, tag or path; give a commit hash or digest only when the check or project rule requires it. Claim no check that did not run; a verdict reports a command you ran. If layout needs more detail, read presentation.

Draw and deliver

With outcomebound on PATH, draw each brief this section's rules call for with outcomebound brief -, JSON on stdin, no file written, using the session surface's marks and diagram form; use --form mermaid where needed. Paste the complete render as markdown in the user-facing text, never a code block; revise input and redraw, never shorten or rewrite the render. If a draw is refused, fix the brief and redraw, never hand-write it. Without the command on PATH, write the same shape by hand with lettered options.

The document is {"briefs": [...], "order": [[before, after], ...]}; every string is one line. Each brief holds id (letters, digits, hyphens), heading (the question), why_now, context (strings or lists of strings, none starting with a markdown mark), terms ([term, meaning] pairs), options (objects with way, leads_to, choose_if, downside), recommend (why), recommend_way (the words of a way, never its letter), at least one of checked ({"verdicts": ["PASS"], "text": ...}), not_checked or facts, and undo ({"can_be_undone": true, "text": ...}); add diagram, picture and explore where they help. Only when a draw is refused, read outcomebound brief --help for full field rules or surfaces needing --form mermaid.

Put several decisions in one message; order draws dependencies. If the harness asks through a question tool with choices, read question tools.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.