Appearance
Task Anchor
Task Anchor is a skill that holds an agent to the original ask across a long or multi-step run. The task line gets buried under tool output as a session grows, and buried context is where attention is weakest, so the agent doesn't forget the job, it just stops weighing it against everything it does next. The skill's fix is to settle the ask once, write it down somewhere durable as a checkable anchor, and hold every action against that anchor until the work is declared done against real output rather than memory. It enforces nothing on its own: there's no hook or gate behind it, it's instructions the agent holds itself to, so a session that never loads it is simply unanchored and nothing will tell you that.
In Practice
It arms off how you phrase the ask. No slash command, no config. Say "stay on task", "keep on track", "dont stray", "stick to the scope", "focus", or "make sure you understand first", or just name it up front on a big refactor, a debugging hunt, or a long tool heavy run, and the skill takes over from there.
Clarify before anchoring, not after. An anchor holds you to a reading of the ask, so a wrong first reading gets held just as firmly as a right one, which is the one drift an anchor alone makes worse. Before anything gets written down, the skill names what's genuinely ambiguous, states what it's assuming, and picks a reading. This runs once at task start, not per turn, and a clear one-line ask clears it in seconds. On a non-interactive run where nobody can answer, it states the reading it took as an assumption inside the anchor itself and carries on, so if the output solved the wrong problem, that line is the first thing to check.
The anchor gets written to the todo tool or a plan file, never left as prose in a reply, in this shape:
task: <one line in the user's own words>
done when: <2 to 5 checkable criteria>
out of scope: <the tempting neighbours>
budget: <rough tool call and file touch allowance>Drift words get parked, not acted on. "While I'm here", "might as well", "also noticed", "quick fix" – any of those is the cue to write one line into the parking lot instead of following the impulse. Found bugs, cleanup ideas, unrelated breakage all go the same way: noted, never fixed unprompted, reported at the end.
Three actions with no criterion progress trips the tripwire. The agent stops, re-reads the anchor, and either picks the next criterion or reports the blocker, rather than improvising a new goal to rescue a run that's stalled or a step that failed.
A mid-task redirect updates the anchor first. If the user changes the ask partway through, the anchor follows the user, not the plan already in flight, and after a compaction or a long tangent the anchor gets re-read before anything else happens.
Declaring done means checking, not remembering. Every criterion gets checked against real output, a test run, a diff, file contents, and every change in the diff has to map to a criterion or it gets reverted or flagged as drift that shipped. The report ends with the parking lot so nothing found along the way is lost.
For non-interactive runs, paste the compact version instead of the whole skill. references/compact-inject.md is roughly 250 tokens, meant for --append-system-prompt or a subagent prompt where loading the full skill costs too much.
For always-on anchoring with no phrasing needed, paste the rules block. references/rules-block.md is roughly 80 tokens, meant to sit permanently in a project's CLAUDE.md or the global ~/.claude/RULES.md so every multi-step task gets anchored without anyone having to say a trigger phrase.
Key Files
SKILL.md
The skill itself – why agents drift, the clarify-first ordering, the anchor's four fields, the while-working rules (criterion-per-action, parking lot, tripwire, redirect handling), and the done-check. Everything else in the module is a smaller extract of what this file already says in full.
references/compact-inject.md
The print-mode and subagent paste-in – around 250 tokens, condenses the whole skill into a flat instruction block for --append-system-prompt or a subagent prompt, including the non-interactive framing (deliver the partial result plus the parking lot on a budget breach or hard blocker, rather than waiting for input that can't come).
references/rules-block.md
The always-on paste-in – around 80 tokens, the four-bullet version meant to live permanently in a CLAUDE.md or ~/.claude/RULES.md so anchoring applies to every multi-step task without a trigger phrase.
When Not to Use It
The skill's own description is explicit about its limit: it is not for single-shot questions or tasks that finish in one or two actions, the clarify pass and the anchor write cost more than a task that short gets back. It's also worth being clear about what it doesn't cover once it is running:
- The anchor holds you to the goal, not the route. Grinding one failing method forever is a different failure, that's what its sibling anti-fixation watches for.
- The two paste-ins are copies, not links. Updating the module doesn't touch a fragment already pasted into a
CLAUDE.mdor a system prompt, it has to be re-pasted after a change. - Nothing checks that the skill actually loaded. If a session never triggers it, there's no anchor and no signal that one is missing.
Connects to
- Siblings with anti-fixation under the same
adherencemodule: task-anchor owns the goal, anti-fixation owns the route. Anti-fixation's nudge and escalation both re-read the anchor's done criteria before acting, so a re-frame it suggests routes toward the same criterion rather than a different problem, and it never overrides a goal task-anchor already settled. See Anti-fixation. - Careful's behavioral fragment defers to it by name for scope and goal rather than restating the rule, and treats its own principle 1 (think before coding) as pre-flight, not a second anchor.
- Declared as a part of the
adherencemodule inharness/modules/adherence/module.json, installed by the installer to.claude/skills/task-anchor. - The condensed, end-user version of this page is the module's own
USAGE.md, atharness/modules/adherence/task-anchor/USAGE.md– that one is the thin command and gotcha reference, this page is the fuller depth of the same module.