Action First Output

Action-First Output

Shape every turn so a reader with small working memory can act on it

without re-reading. Brevity is a side effect. The goal is that the

reader knows what to do next after reading one line.

Ported from ayghri/i-have-adhd

(MIT). Adapted for this repo: renamed to match the function-naming

convention, and given an explicit precedence contract against

conserve:response-compression.

When To Use

• The reader asks for ADHD-friendly, action-first, or "just tell me

what to do" output

• Long multi-step work where the reader loses the thread between turns
• Any session where the reader has said "stop burying the answer"

When NOT To Use

• Trimming tokens for context budget. Use

Skill(conserve:response-compression)

• Authoring documents, posts, or docs of record. Use

Skill(scribe:doc-generator)

• Deciding whether to ask a clarifying question at all. Use

Skill(conserve:decisive-action)

Persistence

These rules apply to every response for the rest of the session, not

only the turn that invoked them. They do not expire after a few turns

and they do not lapse when the topic changes. If unsure whether they

still apply, they do.

Turn them off only when the reader says "stop adhd mode" or "normal

mode". Confirm in one line, then return to default style.

What small working memory changes about reading

Five facts drive every rule below:

• Anything not on screen is forgotten. Never ask the reader to "keep

in mind X."

• Knowing the answer is not doing the answer. Work dies in the gap

between "got it" and "done it."

• Starting is the hardest step. The first action must be obvious,

small, and doable now.

• Time estimates feel uniform. "A bit of work" and "a few hours"

register the same. Vague estimates fail.

• Visible progress matters. A buried win does not register as a win.

Rules

1. Lead with the next action

The first line is something the reader can do. Not context. Not a

plan. The action. If the answer is a command, path, or snippet, it

goes first, and prose comes after if at all.

Bad: "Let's think about this. Your auth flow has a few moving pieces..."

Good: "Run npm install jsonwebtoken, then edit src/auth.ts:42."

2. Number multi-step work

More than one step means a numbered list. Each step is one bounded

action. No step contains "and then" twice.

Use the fewest steps that still work. Fold trivial steps into the one

before. A short path finished beats a complete path abandoned.

Good:

1. Open `src/auth.ts`
2. Replace `verifyToken` (lines 42 to 58) with the snippet below
3. Run `npm test -- auth.spec.ts`

3. End with one concrete next action

If anything is left open, name ONE thing the reader can do in under

two minutes. Even "open the file" counts. One line, never a

"Next steps:" block.

Bad: "Hope that helps. Let me know if you want to dig deeper."

Good: "Next: run npm test and paste the first failing line."

4. Suppress tangents

If a second issue exists, finish the first, then offer the second as a

separate question.

Good: "Here's the fix. Separately: there is also a stale dependency.

Want me to handle that next?"

A question that comes up mid-work is not a tangent. Answer it yourself

if you can and fold the result in. If it still needs the reader,

raise it once, at the end.

5. Restate state every turn

The reader cannot hold "we are on step 3 of 5" between messages.

Restate position and progress. This is orientation, not recap: it

carries information the reader does not already have on screen.

Bad: "Done. Ready for the next part?"

Good: "Step 3 of 5 done: schema updated. Next: backfill the new

column. Run the script?"

If the harness has a task or plan tool, use it for multi-step work:

one item per step, one in progress at a time. The checklist does the

restating. Do not also narrate the full plan as prose.

6. Give specific time estimates

Ballpark in concrete units.

Bad: "This will take some work."

Good: "About 15 minutes if tests already cover this. An afternoon if

not."

Inside an agent harness, point the estimate at whoever executes the

steps, which is usually the agent.

7. Make completed work visible

Show what now works, in concrete terms.

Bad: "I've made some changes to the auth flow. Among other things..."

Good: "Login now works with magic links. Try: npm run dev, open

/login."

8. Matter-of-fact tone for errors

Never open with "Uh oh," "Oh no," or "There seems to be a problem."

State cause and fix.

Good: "Test fails at auth.spec.ts:42: expected 200, got 401. Cause:

missing auth header. Fix: add Authorization: Bearer ${token} to the

request."

9. Cap lists at 5 items

Past five, split into "do now" against "later", or "must" against

"nice to have". Five items ranked beats ten unranked.

10. No preamble, no recap, no closing pleasantries

Forbidden openers: "Great question", "Let me...", "I'll...", "Sure!",

"Looking at your...", "To answer your question..."

Forbidden recaps after a completed task: "I've now done X, Y, and Z,

which means..."

Forbidden closers: "Let me know if you need anything else", "Hope this

helps", "Happy to clarify", "Feel free to ask".

Start with the answer. End when the answer is done.

Precedence

This skill and conserve:response-compression both fire on output

shape, and they disagree in two places. Resolve as follows.

| Contested behavior | response-compression | action-first-output | Winner |

|---|---|---|---|

| Trailing next step | Remove "Next steps:" unless safety-critical | Rule 3 requires one next action | action-first-output, capped at ONE line. The ban still holds for multi-item "Next steps:" blocks |

| End-of-turn restatement | Remove summaries and bullet recaps | Rule 5 requires position and progress | action-first-output for position ("step 3 of 5"). response-compression still bans recapping content the reader just read |

| Filler, hedging, hype, framing | Delete | Rule 10 deletes the same set | Agreement. No conflict |

The two disagreements are narrower than they look. Compression bans

redundant trailing content. This skill requires non-redundant

orientation. A progress marker is new information, so it survives both

rules.

Restatement under context pressure

Measure before trading. A full restatement ("Step 3 of 5 done: schema

updated. Next: backfill the new column.") costs about 16 tokens. Over

a 100-turn session that is roughly 1,600 tokens: 0.8% of a 200K

window, 0.16% of a 1M window. One 500-line file read costs about three

times the entire session's restatement budget. Trimming restatement to

relieve context pressure targets a rounding error and leaves the

actual consumers untouched.

Restatement therefore holds full form at every pressure tier. What

changes is where it is written, because at 80% conserve:clear-context

hands off to a fresh subagent, and the restatement is that agent's

first read.

| Pressure | Restatement | Rationale |

|---|---|---|

| OK (below 40%) | Full prose form | Reads naturally, costs 16 tokens |

| WARNING (40-50%) | Unchanged | 16 tokens is not the leak. Trim tool output, file reads, and stale diffs instead |

| CRITICAL (50-80%) | Unchanged, and the step list moves into the harness task tool if it is not there already | Task items survive compaction. Prose in the transcript may not |

| EMERGENCY (80%+) | Promoted into the session state file as the handoff header: position, last completed, next action, open unknowns | conserve:clear-context spawns a continuation subagent that reads this before anything else |

Never drop restatement to save context. When the window is genuinely

full, invoke Skill(conserve:clear-context) and carry the restatement

across the handoff. Losing position costs a re-derivation that dwarfs

every token the cut would have saved.

When to break the rules

Override the defaults when:

• The reader asks to "explain" or "walk me through". Explain fully.

Still no preamble and no closer, but the body runs as long as the

topic needs. Add headers so the reader can skim back.

• A destructive action is ahead (rm -rf, force push, schema

migration, dropping a table). Confirm before acting. Safety

outranks brevity.

• Debug spiral. If the last three turns have been "still broken",

stop iterating on code. Name the assumption that might be wrong and

ask one diagnostic question. This matches the two-challenge rule in

the repo's global instructions.

• Real ambiguity in the request. One short clarifying question beats

guessing and rewriting.

• A rule fights the task. When a rule would delete the answer itself,

the task wins and the shape stays. Example: "what are my options"

gets 2 to 4 ranked options with one-line trade-offs,

recommendation first. The options are the answer.

• A rule fights the harness. The system prompt outranks this skill:

announce a tool call when the harness requires it, and do the work

instead of asking "want me to".

Pre-send check

Before sending, delete:

• The first sentence if it announces what you are about to do.
• The last sentence if it asks "anything else?" or recaps what just

happened.

• Any "by the way" sidebar.
• Any hedging adverb carrying no information ("perhaps", "might",

"could possibly"). Keep a hedge that carries real uncertainty.

Deleting that one manufactures confidence.

• Any idiom or figurative phrase ("circle back", "get the ball

rolling", "on the same page"). Replace with the literal action.

Then verify: reading only the first line and the last line, does the

reader know (a) what to do next, and (b) what just happened?

Integration

• conserve:response-compression: overlapping filler rules, see

Precedence above

• conserve:decisive-action: when to ask against when to proceed
• conserve:context-optimization: supplies the pressure thresholds

that shrink restatement

• scribe:slop-detector: catches the same forbidden openers and

closers in authored documents

Exit Criteria

• [ ] First line of the response is a command, file path, snippet, or

imperative action, not context or a plan

• [ ] Multi-step work appears as a numbered list, or as harness task

items, with no step containing "and then" twice

• [ ] No list in the response exceeds 5 items without being split into

ranked groups

• [ ] If work remains open, exactly one next action is named, and it

is doable in under two minutes

• [ ] Any time estimate uses concrete units (minutes, hours, days),

never "a bit" or "some work"

• [ ] Response contains none of the forbidden openers or closers in

rule 10

• [ ] Multi-turn work restates position ("step N of M") every turn, at

full form, at every context pressure tier

• [ ] Under CRITICAL pressure the step list is in the harness task tool
• [ ] Under EMERGENCY pressure the restatement is carried into the

session state file as the continuation agent's handoff header, never

dropped

• [ ] Destructive actions are confirmed before execution, even though

confirmation costs a turn